# Nobody buys a point cloud: Florent Poux on AI that turns scans into IFC - Show: 2D Is Against My Religion (2DIAMR), host Michal Gula - Guests: Florent Poux, Point cloud researcher, founder of the 3D Geodata Academy, Neurones 3D - Published: 16 Oct 2026 · 1 h 19 min · Podcast · recorded LIVE - Topics: Spatial AI, BIM, Reality Capture, Digital Twins - Web page: https://www.2diamr.com/watch/ai-point-cloud-segmentation-florent-poux/ - YouTube: https://www.youtube.com/watch?v=_FOWUWnVBwE - Spotify: https://open.spotify.com/episode/3AqSP25pgwYP7lsOUJgcT3 - Apple Podcasts: https://podcasts.apple.com/us/podcast/ai-point-cloud-segmentation-scan-to-ifc-in-seconds/id1878399775?i=1000796652923 ## Summary Point cloud researcher Florent Poux, founder of the 3D Geodata Academy, tells Michal Gula why nobody buys a point cloud – the value is in the refined deliverable. In a live demo of the Smart Point Cloud, Florent segments rooms, answers questions in plain language and generates floor plans and an IFC model, then discusses precision and recall, AI checking AI, vibe coding, open source and copycats. ## Key ideas - Per Florent Poux, raw point cloud data is like unrefined oil: valuable inside the process, but what the end user acts on is a refined deliverable – often a 2D plan that strips away everything not needed. - Manual segmentation is still in massive demand, Florent says, because working at segment level instead of point level speeds up every later step of the pipeline. - Florent advises judging AI segmentation by precision and recall against the final deliverable rather than chasing a perfect classification, which in Florent’s experience takes at least six months of work. - In the live demo, the Smart Point Cloud segments rooms, infers floor and wall areas hidden behind furniture and, per Florent, turns first-generation Matterport data into an IFC model in seconds – cultural heritage (HBIM) is where it still struggles. - Florent compares checking AI output to scientific peer review – never let a model check its own work – and warns that today’s AI token costs won’t last, so teams should own their tools and models. ## Questions answered ### Is manual point cloud segmentation still in demand? Yes – Florent Poux says there is massive demand, also visible among 3D Geodata Academy students, because segmentation lets you work at segment level instead of point level and speeds up every later step of a pipeline. (discussed at 5:08) ### How do you measure the accuracy of AI point cloud segmentation? Florent Poux advises checking precision (is a predicted chair really a chair) and recall (how many of the chairs are found), keeping both in balance and tying the metrics to the final deliverable, such as an IFC or OpenUSD model. (discussed at 13:38) ### Can AI generate an IFC model from a point cloud automatically? In the live demo, Florent Poux turns first-generation Matterport data into an IFC model in seconds; per Florent, the approach still struggles with cultural heritage (HBIM), where shapes differ from regular buildings. (discussed at 36:29) ### Should one AI model check another AI model’s work? Florent Poux compares it to scientific peer review – never check your own work – but warns that LLMs share similar architectures, so for QC you must be sure the checking AI is the right one. (discussed at 43:44) ### Does point cloud segmentation need RGB colours? Per Florent Poux, the Smart Point Cloud works on geometry alone if that is all there is, but uses intensity, RGB, 360° images and any other available data when present. (discussed at 65:27) ## Quotes > “The point cloud itself is not valuable for the end user, but it's valuable very much within this process.” – Florent Poux (2:16) > “This is a scientific principle of peer review. You never check the work yourself. You always have something else that check your work.” – Florent Poux (43:44) > “We tend to forget with hype that fundamentals are everything.” – Florent Poux (73:56) ## Chapters - 0:00 Highlights - 0:59 Nobody buys a point cloud - 4:10 Is manual segmentation dead? - 6:42 Mobile LiDAR, TLS or photogrammetry data - 9:00 Open source models for niche classes - 11:50 The last 5%: precision and recall - 15:18 Never rely on one AI path - 17:24 What clients actually want - 19:32 Solar panels from open LiDAR - 20:43 Demo: making a point cloud smart - 25:38 Rentable area without redrawing - 28:49 How much paint do I need? - 30:21 Walls hidden behind furniture - 32:09 Video to 3D, floor plans, accessibility audit - 33:36 What a simple floor plan includes - 35:44 Where does it break? IFC from Matterport - 39:20 Can a second AI check the first? - 44:31 Education, consulting, development - 46:50 Patents, open source and copycats - 48:16 From land surveyor to 3D deep learning - 50:18 Vibe coding: an uneasy feeling - 53:47 Can code still be protected? - 1:02:15 Favourite dataset: a forest, tree by tree - 1:05:13 Geometry only, or RGB too? - 1:06:40 Is a smart point cloud a digital twin? - 1:09:13 How to try it - 1:13:05 Keeping up when the tools change - 1:17:06 Own your AI before token prices rise ## Transcript _Automatic transcription (ElevenLabs Scribe), lightly edited; times follow the audio._ **Michal Gula** (1:00): you said that nobody's buying a point cloud. Why do you say so? **Florent Poux** (1:10): it's because it's, like if we take a very simple, we translate that very simply, with, with, things we can, picture. We need refining. It's like you, you, you don't buy, at the tanking station, it's refined oil, and so point cloud data is essentially very raw. You have geometry. usually it's very discrete, so it's XYZ, and most of the deliverables or the thing that we act on needs these extra layers that we need to build on top of the data. So that can come fully manually or automatically or semi-automatically. The, the, the path vary, but down the line we really need something refined. So, for a land surveyor, the point cloud will be the working base to create a map that will have different, properties, but essentially there is all the expertise that goes into filtering out what makes sense, what to add on top, and what is value to other people. So the point cloud itself is not valuable for the end, user, but it's valuable very much within this process. but, but we need to go through this process and establish it. **Michal Gula** (2:30): I'm guessing that, this narration you, you started to do this narration once you were able to work with AI and with, some segmentation and, with, with the ability to give some other semantics and, and, purpose to, to those point clouds. Because before it was only about points, geo-georeference, not georeference, but it was only about the measurements, right? **Florent Poux** (2:58): Yeah. But we had,.. Exactly. That's the starting point. But, but we still had this layer of meaning on top of a point. We knew this point was part of a curb. This, we knew this point was represented a, a pole. We knew this point because on site you acquire data, and you give meaning to it by code tables or whatever. But we still attach meaning, and then in the drawing phase, you essentially will give a visual aspect to each of these meaning through layers, through, parametric modeling, whichever techniques we use to, to represent stuff. But, the, the paradigm didn't change. It's the way to get there that has changed. When before we were more, more relying on, completion of algorithm and rule-based mechanisms that were really relying on expert knowledge, today you can more or less prompt some kind of, black box AI that, someone else own, will very smartly take away all that you ask a- and return to you something valuable so that you keep using it. But, but the, the, the results hopefully is kind of very similar. **Michal Gula** (4:11): do you-- We will be showing some, some of your workflows and, and data, but, like generally speaking, do you still see a big demand for manual segmenting of the point clouds? Because I could,.. I'm doing, LinkedIn hardcore almost five years now, and from the beginning, I have seen like many, many, many, many Indian guys, just, just, offering the, the service for scan-to-BIM. and f- from sometime it was also for the segmentations, like, "We can segment your data. We can take along everything you can see on the point clouds." But then it stopped, and I'm just wondering whether the overall manual segmentation is just gone or we still need guys that are verifying and segmenting data manually. **Florent Poux** (5:07): Yeah. Very good. so no, there, there is a massive demand there. and I see that also through the student that I, have at the academy. the thing is that it's a middle, let's say middle step or middle processing stage that is essential to everything, because if we talk about pure segmentation, so just clustering or making sense of, geometrically grouping points that share some similarity criterion, this is very often extremely useful even, if we look at a low-level view where you need to optimize a workflow. So because instead of working at the point level, you work at the segment level, and you cut time by 100 or so. or you want to have more spatial continuity, so instead of having a point-based classification, you will have a segment-based where you have already some more spatial continuity between your object, natural borders or so. So, because it's very hidden behind the actual deliverable, if we take any process, if you have before any process segmentation and, and classification, instead of playing only on geometry you have geometry and semantics, and then the goal is to go much, much beyond to have something extremely strong to cut, essentially the bulk of the pipeline. But at least if you have, geometry and semantics, you are ahead of a lot of people and, and then, and then all your pipelines go much quicker. **Michal Gula** (6:44): can you generally say whether it's, it's more difficult to segment, data from, from a LiDAR or from a mobile device compared to TLS scanners? Or- **Florent Poux** (6:57): Um-... **Michal Gula** (6:58): for example, photogrammetry? **Florent Poux** (7:01): Yes. So, so complexity-wise, I don't know if that would be the, the proper word. Essentially, what usually I, I try to scope is, and, and with, clients I work with as well, is depending on the state of maturity they are in, we'll take a specific approach. So it can be, unsupervised, semi-supervised, or supervised. So if this in clarification is unsupervised, we essentially have data. We know the results ourselves because we're expert. Okay, we have a point cloud of an indoor scan, and we want to find, the floor, the ceilings, the, the walls, the chairs, and so on. And to get there, we don't have at all any training data, so what we do is we go into unsupervised segmentation with, efficient algorithm to do 80% of the work. and then we are smart that we build a human-in-the-loop component that we can take that and just say, "This is good, this is bad. We refine that." And that will create our supervised, starter supervised data set to create a, a 3D learning model or something like it. so this is essentially the path, and the sensory input is a, a bit,.. This is something that we just treat as a pre-processing layer, where we will say, "Okay, it comes from a mobile laser scanning data, so we should expect lower density, et cetera. If it's related to scanners, high density, maybe we have some occlusions." So th-this is just a, a matter of having what we call an ontology that will define, the data rules of where the sensor comes in. But this is something we establish once, and doesn't, doesn't move the needle as is it more complex or less complex. but, but this is only if you have a very strong, smart, point cloud structure or something where your data is really, centralized with some kind of generalizable core. I don't know **Michal Gula** (9:02): if it's clear. So, of course, there are some, some software, some solutions that offer this, like automatic segmentation of, of your data sets. But what if there are people that want to segment something specific from their data because they are doing like very niche scanning, and, they want to train their own data? Are there like open source, models that, that they can use, or how should they work with that? **Florent Poux** (9:33): Yeah. So, for the past, it's more than 15 years now that I'm in R&D and especially on point clouds. So, most of the research that I publish is open source, and a lot tackle these kind of problematics. So of course, they can pick up a GitHub. I made sure also to have open tutorials that's more hands-on and let's, say less research-oriented, but nevertheless can help you achieve what you want. The thing is, today you have to decide of an approach essentially strategically is, is your decomposition of your scene, of the very subject you want something that always happen for all your data that you process? If the answer is yes to that, essentially we are in a closed source vocabulary. So you know you will have 10 classes, and you will predict against that. And in that case, the best is to try to collect lot of data sets for that and, create a, a 3D learning model on top of it. if you are always faced with new object or things that appear, we go with open vocabulary approaches, and this will help you refine and get access to objects that you did not, precedently define, in your list of, supervised approaches. So both are feasible. Of course, the open vocabulary tend to be, much more flexible but a bit less accurate or efficient. But it's, it's something that I use pretty much every day still because, in a lot of cases, there are some small components where you don't want to have that part of the classifier. So, for example, I can show that later, but if you have a table, you have the floor and so on, and on the table you have monitors, you have things like this. And of course, it's not part of the vocabulary that I have, but I'm still able to leverage, VLM's to pick up monitors, have the link to the 3D data, and then be able to, to, to pick up the, these new classes. So the, the, the two are possible. It's just a matter of making sure that the low occurrence, classes, you move them away from the closed vocabulary and you, you use it just with the open side of things, yeah. **Michal Gula** (11:51): The standard bias that I could see is that, okay, I will let AI or any algorithm segment my data. It's claiming, I don't know, like 95% of accuracy, let's say. What about less, the, the rest with the 5%? Like, isn't it gonna take me more time to classify and find those compared to if I was just manually g- doing everything? **Florent Poux** (12:19): Yes. So, that's exactly the points usually. Um- But it's about scoping and making sure the metrics align. So whenever, like if I take a case, at the moment with BMW, the-we, we have a project, cannot share too much details, but essentially there is a need then to define metrics, very clear metrics. And usually the metrics that really matters at the end is based on the deliverable. So if we take an example, that the deliverable is a 3D open USD or IFC model, you want metrics to be able to reflect that so that you can compare against it. And the classification of semantic segmentation metric don't need to reach 95%. You, you need to, to have the other part of the pipeline flexible enough so that it can cope with, what we call label noise, so error in classification. because wanting a perfect classification is, is really an effort, that you cannot have below six month of work. Now, if we take into account AI, maybe you can lower this barrier, but usually you really need to have a lot of data but very good quality. and your 3D deep learning model to be able to generalize would need to have a lot of, engineering work in it. So that's what I can say. Now, as for the metrics themselves, what I advise is to check two metrics for most of the scenario it works properly. You have a precision. You predict one thing, it's these things. You predict a chair, it's a chair. and you want to check that against the performance of your model. Does it predict chair very precisely or sometime it's wrong? That's the precision. And you have the recall, which is, if we have, 100 chairs in the room, how many do we pick up? Only 90 or 95 or so. And usually you need to have a balance of both precision and recall. So exhaustivity, you want to find all the object on the line and be precise about them. So that's what to emphasize for. But this is totally linked with the use case down the line that, this is what we, we, we need to feed in, especially if you use AI for coding, this is what you need to feed in your, your architecture. **Michal Gula** (14:32): Okay. Th-this is quite important point because, yeah, not everybody is realizing that, th-these two variables can, have different effects on, on, on the, the final results. Because one thing is that whether the, the model is able to recognize like or how many percent of, of the, of the items is it's able to recognize. But if the percentage of, let's call it false positive or, or you were saying of recalling the, the, the right ones. If, for example, I need to pick up chairs, but from those 95 d- diagnosed chairs, there are like 10% of tables, not actually chairs, this is even worse, right? Right. **Florent Poux** (15:18): Yeah. Yeah, yeah. And this is why I can also relate to, recent work. So, for example, on, there are some specific system that are very sensitive to, like stamping quality control and making sure that you get the answer not from one way, but multiple way. And this kind of system are usually critical. So of course, with, BIM modeling also it's not critical. you, you can recover from it. But if we talk about, autopilot system, for making plane lands or for trains, usually operates, especially for planes at ten minus nine. So you want one casualty every, billion, flight. so, so, so one... So, so of course you operate with, you have to imagine that if you were to delegate that to AI, you need ninety-nine point nine nine nine nine nine nine nine, and that's none do that. So usually what happens is you, you have a lot of recovery mechanism, and you have late fusion. So if you have a, a default in one path, the two others should be able to carry on the task for you. and this is something that needs to be put in place whenever it's critical. The accuracy is that you don't arrive to a metric only from one way. And especially today, I think it's worth emphasizing, it's very dangerous to rely solely on AI to, have something against this. When I say AI is the big player at the moment, so Opus, from Claude and, Astra from GPTs that works a bit with, 3D and all the new from Google that they released today or something. this is one way of doing things, but of course, it's, it's best if against it you have different, mechanism or d-d-different path to get to the same result, and then you can do late fusion. And those are the kind of thing that are done a lot when you have multi-model data acquisition, and we do. So, if you can leverage that, that's, that's a great, starting point, yeah. **Michal Gula** (17:25): we started with the question about why you're saying that, clients don't want point clouds. So before we jump into actual showcase, can you tell me what they actually want? **Florent Poux** (17:41): Cannot speak for all clients, right? So it's, it just a matter of, of, conceptualizing things. But for sure, i-if, we take ourself or people that I work with, usually, especially in Europe or in France, we're really attached to, to 2D deliverables, let's say. Sometimes 3D, but it's, it's not, it's not often the case. And with 2D deliverables, really the value, lies in what you strip away. So essentially, I don't know who, who created that, but a very long time ago, there was this guy that, created this book where you had a one-to-one map, and it was totally useless. Like, and that's what we do with splits. We scan everything, but it's absolutely useless because what we need to act on is something where you strip away what is not necessary, and you only keep what you need to take action on. So this is where you go in a cycle of, what I call a capture, understand, deliver, and act on. And this deliverable component really is a way to present information and present everything in a data structure that is good enough so that you can act on. And this action is either human or delegated. That's up to you. But, very often, if we think about humans, it's a 2D map. It can be a, a, a map to, to see the legal boundaries, it can be an evacuation plan, it can be auditing for a wheelchair, it can be a solar estimation of, of an area, urban cadastre, you have a robot mapping as well. All these kind of things, we tend to work really well with 2D representations. so yeah, th-this is usually the line, that I call scan to deliverable and scan to map. **Michal Gula** (19:33): So is, is it fair to say that they generally don't need information, but their goal is to make a decision about something? **Florent Poux** (19:42): Yeah. Yeah, yeah. So a-and even, so for example, being a bit crazy, I recently put solar panel on my roof, and I decided not to go on my roof. So what I did is I went and got the open LiDAR, and I have, with the smart point cloud that we show, I have a mechanism that will model totally, extract the roof surface and other solar potential. but all of this is for me to be able to act on it. So after ordering the solar panel, and making sure it's well dimensioned and pushing through all the administrative components to make sure that I have the right to do that. the, the base of all of that is raw point cloud. and it's just a matter of refining that so that I have enough material to act and actually make it happen. So of course, no robot went and installed it. I went with a friend and, and we installed it, but maybe in the future we'll have robot that can install that. Who knows? But th-this is the thing. We, we, we can act on something that has enough, enough components to act on. **Michal Gula** (20:44): I think it's the right time to, to show something to our audience. So, can you practically show us what you mean by making a point cloud not dumb, but actually smart? **Florent Poux** (21:02): Yes, sure. So you have various components that I can show, so don't hesitate to interrupt or ask questions as we move along. So this is, for example, Neurones 3D Science. This is, what I call a bench of experimenting where- I pushed every algorithm I ever developed, every AI models, including classifiers. every workflow is there so that you have kind of an SDK, but it's open, so it's all Python, and all the students in the academy, they have access to that, and they build a lot of thing on top of it. So it's cut down a lot of costs, and you are sure about, the, the reliability of what, what is behind, let's say. So what I meant is if I open it, the cycle with this capture cycle where we have the, the point cloud or splats, whichever you prefer to work with, the bigger point is this understand, layer where we need to make sure that the point become knowledge. So we have the segment, the classes, the rooms, if we take a, a simple example, and we can extend, and then we have this deliver. And what you see on the right is something today is one click away. It's-- I'm a scientist by, by practice, so I'm very careful about, this one-click stuff, but here we are at, at this level. And then after you act on it, and if you do re-renovation, you order the paint. Accessibility will widen the narrow doors if you want to inspect, if a wall is out of plumb or, or all these kind of things. So now if I take this point cloud here, which is of, of an indoor, indoor building, the several layers that we'll have are the following. Usually, we have to go in the smart point cloud infrastructure to bring more than just geometry. So we'll get with unsupervised approach. We will have a classification. This is from, one model that I built that is called Lucid. **Michal Gula** (22:54): so sorry. Yeah. what you were showing were, different rooms labeled with, different colors. What was this segmentation done by? **Florent Poux** (23:05): So this is, unsupervised segmentation, and, early on in twenty nineteen, so I have the reference here. In twenty nineteen and twenty twenty-two. In twenty nineteen, I published a, a paper that won the Jack Dangermond Award, and it was about benchmarking at the time, a voxel-based feature engineering with deep learning. It was very heavy deep learning at the time, for three-sixty data. And seeing that with unsupervised, workflow, there is, a very, like very powerful capacity to actually generate sound segments, and this is what we can see more or less from here, that will create a, a very good base for the following up, of the processing pipeline. So how they are defined is essentially making sure that we try to respect the boundaries, the natural boundaries of objects, because the limitation is if you have an issue with, segmentation, recovering from it means, reworking a bit on the data structure. So if the segmentation is faulty, then you have an issue. So this is what we have here. And this is then used, to have this classification automatically done, and this is what I talked about. Here, we are in a setting where we are in a fixed, vocabulary classification. We have wall, ceiling, floor, clutter, all of that detected automatically. But we don't stop there. We have semantics, we have segments, and what is important is per class to get the objects. So this is what we have here. You see we have, clutter, all the components, and this to get a sense of space because we can now have, objects, we have the structure, so we can structure our environment with spaces and subspaces. It gets a bit more technical here, but the meta model of the Smart Point Cloud below is able to understand what happens to the, the data at hand, and that's the starting point to create what we call scene graphs. and you see this is of the entire scene. And if I take a specific, room, I don't know, S7, you get the relationship below of all the components, and this is what is missing is in almost all approaches today, is this sense of relationships allows you to extract a lot of contextual knowledge that feeds any AI and that will allow to act on. And- **Michal Gula** (25:36): that- **Florent Poux** (25:37): Yeah. **Michal Gula** (25:38): I, I can see a huge potential here, because I have had several, clients which, they are renting a lot of, spaces, and they are counting everything, down to square meters, like e-even the, the cost of the food for their, for their, for their, employees. the idea is that rarely there is the case that the documentation they have actually fits the reality. So the typical case is that they are renting different amount of square meters compared to what they could be renting or, or compared to, to the reality. So the typical workflow is we scan, we either BIM model or, create 2D drawings, and then calculate the, the square footage and, and segment the, the different r-rooms. With what you are showing here is that we can like generally skip the whole modeling or, or, redrawing and just get the data, right? **Florent Poux** (26:47): Yeah. But it's not incompatible as well. So, for example, this is the next layer once all of this is set up, and this is the most exciting things at the moment that I'm heavily working on is this novelty of, world models. and specifically for three-D. With all the backlog and research and knowledge, I was able to pull out a new world model that I called N three-D Resoner, which is frugal by, by design and, and local. 'Cause, with all the clients I work with, I'm, I'm very, conservative, and I don't want to push any of the data through The big players out there because, yeah, it's, it's something that you need to take or keep close to your chest. The data is really important. so I built this model that is able to do everything from the understand layer to, acting. And I have this various component, like understand. I, I could ask him, "Where are room three and room five?" And it will be able to check on the data and tell me, where they are and show it to me as well. So all this kind of component can be done, in close to real-time. **Michal Gula** (27:58): I was aiming to, to one question with the, with the premise that I said before. are you able to... Be- because, like, the measurements of rentable spaces, this comes with some regulations, with some standards, and with, with some, some, mm, typical ways of how to perceive the space, what, things to count in, what not. to what detail are you able to, to segment it automatically, the square footage or s- or square meters? For example, if, if some standards i- is saying that the, the, middle way in, like, like the door space is not counted into the, into the rented space, like, can you do this automatically, or do you need to redo it manually afterwards? **Florent Poux** (28:50): Yeah, yeah, all is automatic, and it can also... So here I am showing the understanding, like you can ask thing to understand, but not only on the chat. You have it then so visually it, it goes to it, and it show you. And you can go much deeper. your question is one of the best, is the acting. I could ask him, and I will do it because they have, a word to text thing. I will ask him, "Can you please check out room seven and segment it into its main components?" And normally, it should be able to pick up the room, and you see, and segment it in real-time. So those are the kind of thing that are accessible now, and you can go much beyond. So I could ask him to create a BIM model of this room. I could ask him to now generate,.. Yeah, I will do that. "Can you please, check how much wall there is, square meter wise to tell me how much paint I need to order to paint?" All the walls in, white. and normally should understand, let's see, for the specific room. Oh, it took all the rooms. oh, yeah, I didn't say a specific room. So it tells me-... for all the room, I have two thousand four hundred and twenty-five square meta, time two coats, four hundred and forty liter of premium matte wall paint, which is about six thousand euro, or seven thousand euro. **Michal Gula** (30:09): So- **Florent Poux** (30:10): And, and we see it so we can check, all the walls that are selected also. So those **Michal Gula** (30:15): are- Uh-... the kind **Florent Poux** (30:15): of things that the Smart Point Cloud allows you to do, where you plug in this kind of interface, **Michal Gula** (30:22): that- This is g- this is going to be a stupid question, I- I'm sure. Yeah. It, it, it's counting with that, but just to make sure, it, it's not counting this only on a visible points, but it, it is recalculating the whole, like planes from the walls, for example. Because mo- like partially these walls are hidden or floors are hidden, under the furniture or behind the furniture, but it's still counting with those,.. **Florent Poux** (30:53): Yeah, exactly. No, no, that's not a stupid question. Because we have this notion of space, there is the ability to infer that, oh, this is occlusion. all this, thing below, chairs, this is occlusion. This is actually part of the, the, the floor area. But we can check against it as well, because as you know, we work with real data. This is not synthetic data, so we need to cope with that-... else it's useless. not useless because for robotic simulation also you can go with s-synthetic data, but here in this case, it's, we work with real data, so we need to, to handle that. so yeah. A-and of course, in the act, or goals or this kind of things, we can go, and as I said, because behind we have access to, open vocabulary things, I could ask him, "Okay, in any room, what is standing on my desk?" And he will tell you, "Okay, we have, two monitors and so," and then I could go beyond, "Okay, I want to replace this monitor. Can you check the place if I can put one? if I need to put a, a fixture on the, on the ceiling, how should I proceed?" And this whole idea actually started because I, renovated my, my, my, my house and again, crazy enough, let's make it with point cloud. So there is also, you see it here, video to 3D. You just drop a, a video, and you get the full 3D reconstruction of it. and then after you have all that behind the Smart Point Cloud, that allows you to go much beyond. And, and much beyond is, what you see here, the next steps, deliverables. So this is what you generate. This is a simple floor plan, but this is a one click as you see. You have the space and you have the floor plan on the right. the same thing with the BIM model. I could generate the BIM, IFC BIM model here. Accessibility mapping. If I run a, a wheelchair audit, this is a case that I had also, is can this data set be good for people, with, accessibility or condition? So here, of course, we have to check what is okay, what is not okay, and we generate a floor plan. So you this is the, that you will get. It's the PDF with, in green what is good, in red what is not good, and, below you will have all the specific checks that will be found to, okay, here the passage is too narrow, so investigate that. And of course, we have the link to 3D. So this is, for example, for decision maker, something very, very important because it's allows them to make critical decision very quickly. And in emergency scenario, response scenario, this is things where you need a critical, insight in a very time-limited fashion, yeah. **Michal Gula** (33:37): Let's, let's discuss, these Uh- Yeah. And I wanna discuss them separately. so let- **Florent Poux** (33:44): Okay... **Michal Gula** (33:44): let's go to 2D, first- **Florent Poux** (33:46): Okay... **Michal Gula** (33:46): because you said that, "And here is a simple floor plan." what do you mean by that, that it's simple? **Florent Poux** (33:54): I will, open then a s- a case specifically to talk about the, the scan to floor plan. But a simple is, let's not go too much into details about what we present there. So essentially what you saw before is just room, area per room, openings, and that's about it. But i-in some other scenarios, you need also the height above the floor, the height above the floor. You need all the measurements of all the walls, all these kind of things. But there's, extra attributes. All the data is there, just a matter of representation. So in this case, this is another example, which is the, a, a classroom that, you can see here, and on the right side is this room that is generated, automatically. And we could also argue that this is simple, but the extra stage here are that the system is able to put the quotation well aligned. the door is upside down for some reason here. And if we wanted to go a step beyond, we will make the chair a bit nicer than the cuboid. But already here you can see the backrest and the seat and the various table. and these kind of things, because as I mentioned before, the scan to deliverable, usually you will have it, as a DXF or DWG file, so this is what you have there, or as a PDF. And these are the component where, you gain a lot of, of time. If you have in your process the ability to scan data, then y-you can generate this kind of component. Not all is perfect, but for cases where your data is pretty, let's say complete, it does 100% jobs. and you can see the value steps, yeah. **Michal Gula** (35:44): Still this is quite a simple room. Like- Yeah... very square, very simple room. So, uh- Like, I'm the first advocate of, of the AI in our business and, and I, I really think in a few years we'll be generating IFCs and, and DWGs like, like in, in, in minutes from the point clouds. I'm, I'm pretty sure about that. So this is, this is not trying to, undermine your work. I'm just curious, like right now, where does it break? Because you were showing a simple room. So if you are using right now your model on a more complex structure or, or, several floors of the building, and where does it break? Like where do you see limitations right now? **Florent Poux** (36:30): This is another example with, data that is totally incomplete and really not proper, and this is to support your example of IFC generation, and this is what you have here at, at the end. So this is the entry point and, down the line you see the value steps where we will get the slabs, we get the orientations of the walls. We'll be able to go in, various, stages and, get to the drawing So yeah, th-this is a great example because we mentioned it before. You see here, if we take the point cloud originally, this is what it looks like. So we could argue it's simple, yet we have windows, we have columns that are hidden with a plastered wall. Here as well, we have openings that are different and needs to be modeled as such, and we have other components as well. Desks, chairs, everything is touching, and the data is a bit crappy because it's from a Matterport original sensor. So- **Michal Gula** (37:33): No, no offense. No offense to Matterport. **Florent Poux** (37:35): Yeah, yeah, no offense to Matterport because the, the, that was I think one of the first generation. When I say crappy, it's not, just the, the quality of the data as opposed to what you get with the terrestrial laser scanner. It's very, very, uh- It's-... very valuable. **Michal Gula** (37:49): It's a simply objective, objective- **Florent Poux** (37:51): Yeah... **Michal Gula** (37:51): feature of the, of the, of the technology they are using. It's, it's not a bug. **Florent Poux** (37:57): Yeah. Yeah, yeah, exactly. and you see the... So this is a slice, but I could get the IFC model or, or the slice. Let's check the, the slice, to see the kind of plan that you, you have from there. It's pretty exhaustive 'cause you will still see double door, door, door here, also the various components in the space. And if we look at the, the various component, they are well modeled. You see the double door, this door here, the cupboards. I've also attached this, kind of library where you have all the chairs independently. The limits, at this stage is, architectural, not architectural, cultural heritage, where you have shapes that are really different from what we are used to seeing because they have some regularization component in, in the flow of things. But it's not often what, people encounter when they work with, with BIM. It would be more with HBIM that it struggles, and it's another approach that we should use. But with the BIM part of things, yeah, it's in a matter of three seconds you have this entire IFC file essentially. But this is, again, this is just one scan to deliverable, aspect of things. Scan to CAD is another one. there is another one that recently, do I have it here? yeah, **Michal Gula** (39:21): for robots. Have you, have you o- have you already thought about how to do also automatic or, or world model back check? I'm asking this because, ju-just today I, I was working on, on some task, and the output from Claude was, like pretty neat. I, I was almost dead happy that, that I sent it out right away. But then I just, did a test and asked ChatGPT to do a fact checking, to do some, some technical side check, and it actually found some mistakes. It went back to the Claude, and Claude admitted that, "Yeah, this is wrong. We can re-re-- like fix this." So my question is whether you think that, your algorithm or your, model will analyze everything. It will create IFC files, for example. And if you used another model just for checking how it fits, do you think it would be able to, to find mistakes? **Florent Poux** (40:29): And to- Yes, I, I guess so... fix **Michal Gula** (40:30): itself? **Florent Poux** (40:32): yeah, yeah, yeah. There is a self-improving loop, indeed. So, so this is very important. Whenever we build this kind of solution, usually you have a... If you design it properly, you have this self-iterating loop that will improve as we go. and this is something that is there, baked in. But you can change the core of... Like I'm trying to think, like usually when we design this kind of a solution, you, you think modularly, and you think in the future. Like what you said before, you think in the future we'll have this one-click thing. I think the future is, much closer than we anticipate, 'cause this is an example of the one click that will have limitations, but it's about, some tweaking. The thing is, always thinking that what is good now will be obsolete in a matter of now it's very quick. **Michal Gula** (41:24): But, but- But **Florent Poux** (41:26): try, try **Michal Gula** (41:26): to understand my question because- **Florent Poux** (41:28): Okay... **Michal Gula** (41:28): self-improving is one thing. There is this saying that if you wanna master basically anything, you need to spend 10,000 hours, doing it. But there is also another saying that, doing it by yourself and repeating, same mistakes... all over again are not going to, to make you master, right? So you even during these ten thousand, hours, you need somebody else, some mentor, some trainer, some, some coach to tell you what to improve. That's why I'm asking about this, because you have one model to build it, but it should be **Florent Poux** (42:05): a **Michal Gula** (42:05): separate independent model to tell you this is wrong. **Florent Poux** (42:11): I agree. And this **Michal Gula** (42:11): might be very different training, you know? **Florent Poux** (42:14): Yeah, no, but it's the case. It's the case. And there is always, so let me illustrate maybe you, you mean in the loop behavior. if I have a data set somewhere that lies. So this is the other part of it. So, so what we see here is Neurones 3D Science, and this one you can get it for free. It's Newons 3D engine, and I built it in Rust because it's much more efficient than it's, desktop. and yeah, so yeah. Y- we will sit here. So this is another example. So this is the prediction of an early stage model, for example, that, that should predict in a factory various components. And sometime you, you need loops. So it's always important to build tools to create that for yourself. So manually, semi-automatically, and with AI assistance. And this is, the three layers here that we'll have. So I will be able to select the various components here. We see the class. We, we could also model this object, as you can see there. not the, the thing. But if we wanted to refine it, we could, we could by clicking, by giving some kind of feedback, but I can actually record myself and go, and everything is synced with my scene. And after recording and seeing this is wrong, please improve that, this is something where we have a mistake, then the model will be able to pick up and rectify that. So this model, it's up to you what you choose. And I agree with you that... But this is a scientific principle, of peer review. You never check the work yourself. You always have something else that check your work. And in this case, indeed, it can be another model, but you have to understand fundamentally that, other model, they share the same architecture or something very similar, that they are LLM, so they are dumb, and they just use tools, and these tools are algorithm that actually coded and works in a specific, direction. So if you want to do QC with AI, you need to be sure that it's the proper AI that you use to do, QC. But yeah, definitely the principle of, isolating buckets so that you have something that scientifically sound this is, this is true. **Michal Gula** (44:33): you are a You are, basically a developer now. where, where is business? Like what you are showing here, is this only for, for, like learning purposes and, and education and, and research, or is this something that you are actually build- building for, for sale as, as a service, or how does it work? **Florent Poux** (45:00): so yeah, as you mentioned, e-essentially my, my position is, **Michal Gula** (45:06): I **Florent Poux** (45:06): act in three It's very It's education, so we have the 3D Geodata Academy, where my goal is to make sure that everyone is equipped with the knowledge and the tools to build a solution that has value to other people. That's what is most dear to my heart. This is in, in a kind of a D- do it yourself fashion. So you have everything, and I explain it to you, so that you, you master the, the fundamentals so that what you build is sound. The second aspect is consulting, so usually, you need some kind of lead of your R&D team or you, you need to, to make sure everything is progressing as planned or you need a direction, and this is something that I do. And the third is, developing or scoping some kind of development, and this is what I do with, Airbus, BMW, Meta and some other companies. But down the line it always revolve about education, so this is what drives me. Also at the university, of course, ed-educating because, we learn so much. I, I also do because there is all this use case, all this industry, and you cross correlate things that are done somewhere to another one and, the novelty is much, much more impactful there. Now, as part as a software or things like this, the way I see things is, yes, I, I, I could make it. I already got major offer for some components because the IP behind is, is tremendous. But what I search for is really, making sure that the people that gets ownership of this, have the market penetration to, to make it a fantastic product for, for whichever market they, they want to have. So that, that, that's how I position myself today. **Michal Gula** (46:51): do you already have some, some patents for parts of the IP or, or how do you protect yourself? **Florent Poux** (46:59): Yes, I do have patents. I do have also, publications a lot. so, so that's more for academia, but whenever it's published, of course it's, it's either in the public domain or protected somehow. and everything is documented, so all the code, everything is there. In, in coding we still have this very naive approach where it's community driven, right? So you help people, you don't expect anything, but you help people achieve what they want. And hopefully in a world where we are all a positive thinker, this is something that, grows society all together so you don't necessarily need to protect, so deeply the thing. So I guess there are some holes in the racket if someone wants to screw, something over, but, usually it's not the case 'cause especially today, I often got some copycats that took a, a specific component. But then some student of mine or someone in the community told me, "Oh look, this is, word for word what you, what you have in your book, 3D Da- Data Science With Python." So in that case then the editor take charge of that and they do whichever legal action. But yeah, this i- this is really not too much what, what, what drives me. **Michal Gula** (48:17): be-before the Vibe coding was the case, were you a programmer? Like are you a coder by, by heart or, and by education? **Florent Poux** (48:27): yeah, by education. So essentially my first diploma was a land surveying. So in France you then do two years to, to be a land surveyor and at that time it was back in the days we were scanning a lot, creating in AutoCAD floor plans and so on. Two years? Two **Michal Gula** (48:44): years for a land surveyor? **Florent Poux** (48:46): Yes. So you do five years to get your engineering diploma. **Michal Gula** (48:49): Ah, okay. And then **Florent Poux** (48:49): after you work for two years, in practice to do a lot of jobs and then you, you go in front of a jury and you get accredited, if it's good enough. But during this two years working, this is where I was doing a lot of scanning jobs and drone jobs and LIDAR jobs and so on, but always redrawing everything and this is when I switched to solving that programmatically. So I did my PhD on, the topic of point clouds and that was like 100% Python, and C++ and, and 3D deep learning. At the time I was working also with, Stanford a bit, you know, with this PointNet, a kind of approach where we, where we're doing, risk components. So that was early days of 3D deep learning. and then after that through postdocs in geometric deep learning and so on, you refine your art, you get better at coding and so on. And today basically, AI can replace the coding part but all this knowledge, is very useful to, to drive because you know exactly what to search for when you code, what you need to be careful of when you do server side, a client side, the protocols to use or, or all those kind of things. so it's even more fun. The only thing is of course it, it gives a lot more ability to everyone to code, or to expose ideas and this is really what, what is interesting to me- **Michal Gula** (50:17): What's-... seeing this **Florent Poux** (50:18): idea to deploy, component. **Michal Gula** (50:20): Yeah. What was your first feeling when you realized that you can actually save, not days, you can save weeks or months, with coding if, if you use Vibe coding? **Florent Poux** (50:35): I, I was, feeling like it's an uneasy feeling. So if anyone in the audience had, like me, a, a developer background for at least a decade of programming, you know that sometimes you spend one week debugging a loop, a for X in range, blah, blah, blah, do this and this. And for some reason, the thing breaks, and you spend a week debugging that, and you're like, "What? What the hell?" So it's not the fun part. **Michal Gula** (51:05): So- Ju-just, just to find out that there is some, space or- **Florent Poux** (51:09): Yeah, yeah, yeah... or **Michal Gula** (51:09): ju-ju-just a point or whatever. Yeah. **Florent Poux** (51:12): It's usually some crappy things that you forgot because you are so deep in, in something and, and you're not well-organized and whatever. And y-you feel really frustrated, but then you have this aha moment after one week that you find the mistake. You f- you're like, "Yes, I have it." I-- With AI, you don't have that anymore. But, but I do have that with IDs and making that happen because the way I will drive AI is, is never leaving him, like build for me a scan to BIM stuff. It will be knowing exactly which algorithm to use when, and when to loop everything in, and I will ask him early on to just launch a lot of research agent to pick up literature so that I can take a, a fully-fledged decision early on. So it puts me at, like initially I was like, "Oh, crap, I spent 10 years doing, developing for, for, for, yeah, for, for, for nothing." But actually no. And I think the only thing that scares me a bit is the pace of thing, which is very high. And before you know you had a team of developer or a researcher working on a problem, so you had a lot of feedback from various perspectives and a lot of halts possible at, stages, and you were filtering out a lot more bad ideas or bad, components. And today, because everyone can code, there is an amount of information or prototype that is massive, and the, the struggle is then to find the, the proper ones or to, to, to filter through the noise essentially. So, so yeah, that, that's, that's my feeling at the moment, but it changes from day to day because, it's so quick and it's very hard to, to keep pace even on the literature side because now when you are in academia, you review a lot of papers onwards and you, you say, "Okay, this paper, is sound, but the experiments, it's not good," and whatever. and today, all the papers that I review, it's all formatted the same way. It's incremental benchmark improvements, and there is no unique perspective. Or it's very rare that I find a paper like, wow, this researcher or this research team, awesome. it's very little. And the major research happens is three or four companies, now. And this is a, yeah, this monopoly i-is a bit scary, yeah. But yeah, **Michal Gula** (53:49): it's- The, the reason why I was, was asking you whether you are a coder by heart and, and by, by, what you do every day is because a typical problem with, with programmers is that they are afraid about their, knowhow, about their code being stolen because, like the GitHub was, back in the days, it was a holy grail from every, coder to go to, to learn something, to use, to, to build something new from it. But now GitHub is just a second away from basically everybody, not just from coders, because AI can easily go through GitHub and, and build anything you want. the, the bad thing, what's happening is that it can actually steal things from, from those codes, steal some parts of, of the code. So how-- What do you think about this as, as a coder? Because if I'm not a coder and I'm using vibe coding, okay, I don't care. Like if anybody steals this, okay, so be it. But you as the one who was building something for weeks or for months, to know that somebody else can just steal, write this with the AI, **Florent Poux** (55:00): how **Michal Gula** (55:00): do you feel about it? **Florent Poux** (55:03): I mean, my-- Like what I will thrive for is to open everything that I do. So I have no issue with that. I'm very happy, that if everyone can take what I do, really do it, just mention the source. This is the only thing. You mention where it comes from so that there is a, a relatable thing so that we can pick a preference and we can, you know, collaborate with this open-mindedness instead of stealing. Be-because else it's called stealing, right? you steal hours of work of someone else, and you put your name on it. So that's the only thing that I would say. If you can be careful about it, you can experiment as much as you want, but as long as when you want to push that public, just check the licenses of what you have, how you can use it, when you can use it. If you can do that at the early stage, that's even better, but not everyone does that. So no, no, I, I'm very happy if, with this openness. Of course, it means the, there is a collapse in the costs, and the value of everything that is produced information related. but personally, this is not where I position myself. you-- This is more this human side of things, like making sure you get the proper path forward, and then all the rest is tools. So it's really deline-delineating, sorry, what is, methodological conceptual from the implementation essentially. So that's two layers. So I'm not too worried about that. But of course- personally for the business that 3D Geodata give me to, to thrive, I have around 95% of what I produce that is open today, and 5% that is paid. and the way it works is very simple. the paying side of things will help me create more open content, and this is a, a virtuous But yeah, it can, it can happen that so- some people take the courses, and then, get away with the code and create things and,.. But you have guardrails. When you code, when you know how to code, there are ways to actually stamp whatever you do. Everything is breakable, but this is a, a very large and, like cop- copyleft license somehow or, or so, so that it's easy if there is a legal action, it's very easy to prove that's your code or not. **Michal Gula** (57:27): This might be also a time in, in, history when, when in the future we will say just, "Oh, do you remember when code and IP was actually protected and, and actionable?" Because, right now if you want to protect your IP, your code, your, your software, you just need to hire a bunch of lawyers and, and, to fight everybody that tries to steal, it. But it was only like few days ago, I don't know if, if, if you saw this, that, somebody, just like recreated the whole Adobe- **Florent Poux** (58:07): Yeah, exactly. **Michal Gula** (58:08): like software library. That's really crazy. **Florent Poux** (58:11): Yeah, yeah, yeah. my friend, which is like the most amazing developer I know, a full stack developer, created this game for, for years alone in a dark room. A very cool game. and yesterday evening after a tennis session, he told me, "Okay, I used AI, Flo, I, I used AI. I know it's against my nature. I did, and I'm frankly amazed by the quality of the, the code in Rust it could write for me. And, and I'm a bit scared now, but it's, it's amazing what it can do." And then he told me after, the story yesterday It's about Adobe and so on. And indeed today you can refactor or, you know, or, retro engineer everything. there's like this specific GitHub also stuff that allows you to retro engineer. What does it mean? I agree with you. Um- **Michal Gula** (59:12): Yeah, you know, this, this- I don't **Florent Poux** (59:12): think it's... Sorry, yeah, go ahead. This, **Michal Gula** (59:14): this, this really in the future might be just an episode in time. Because, for example, if you, if you take like, money, right? Like back, back in days it was backed by the gold. So you always knew that, for the amount of money that you are using, there is always a portion of gold some- somewhere stored, which really stopped being the case. The, it's, it's like tens of years since, since, this is not the case. And, and the co- not companies, but, like, governments can, print money on the go, how, how much they ever want. And this is, this might be actually the case also with the, with the coding and with protecting of software, like really with new and coming AIs, it will be simply impossible to, to, to protect it. So it will stop being the case. **Florent Poux** (1:00:12): Yeah. This is a bet that you can make. Yeah. Indeed. I don't know. I don't have the answer here. I don't have all the variables. but, yeah, definitely it's going to be harder. But look- E- **Michal Gula** (1:00:22): e- even a-... against, **Florent Poux** (1:00:22): the Adobe history or story, I guess Adobe can do a class action or something against the, the guy that put it open source. I didn't follow it, all the way through, but if it's very similar and if it's infringed **Michal Gula** (1:00:37): Well, he's, claiming that, there is no code used from, from the original. That it was only like, following the, the, the structure behavior, the, the experience and features. But, he's claiming that the whole code was built like from the scratch. Okay. **Florent Poux** (1:00:54): Yeah, yeah. So on the code- Nothing was **Michal Gula** (1:00:55): copied... **Florent Poux** (1:00:55): yeah, you're right. On the code side, maybe, maybe it's harder to, to prove, but on the patent side it's not. Like if the method he uses the one described in the Adobe patent, it will be sweet for that. Yeah. It, it **Michal Gula** (1:01:07): might be like with luxury bags, like it doesn't matter that- Yeah... you didn't use, uh- **Florent Poux** (1:01:12): Yeah... **Michal Gula** (1:01:12): materials from the brand- **Florent Poux** (1:01:14): Yeah. Yeah... that **Michal Gula** (1:01:14): you just recreated it, it's still theft, so But **Florent Poux** (1:01:20): yeah, that's, personally that's not where I, I, I get up in the mornings, so this kind of **Michal Gula** (1:01:25): stuff. No, but, but this is exactly the case because you said that, most of your work is open source. **Florent Poux** (1:01:30): Mm. Yeah. So **Michal Gula** (1:01:32): you are basically protecting yourself, for the future because you simply don't care. **Florent Poux** (1:01:37): Yeah, yeah, yeah. It's, but because I have the luck that I can do that. But if you are in a position where you need to put, food on the table, everyone does. and yeah, today, let's not talk about that. But yeah, if you're really in this early stage, kind of things where you really need your paycheck and this is tied to your entrepreneurship, way, and it is tied to your code and so on, yeah, then, then it's very worrisome. But hopefully this is really the other way around. Big companies will not attack like very tiny creators so on, because that would be worrisome. Yeah. Mm. **Michal Gula** (1:02:16): from what you have tested so far, what you have done, with the, with the segmentation, with 3D model, do you have like your favorite data set that you can show us? **Florent Poux** (1:02:27): do I have a favorite data set? any point cloud that I, I fell in love very early on with point **Michal Gula** (1:02:34): cloud. No, no, no, but, but y- you know, something that if, if, let's say you meet a pretty girl or, or, or, somebody you wanna just, just, brag in front of and you wanna show something interesting, what would you show them? **Florent Poux** (1:02:52): a very good question. So I have one thing, but, maybe I hold that for later. It's, it's amazing, but le- let's hold that for later. I have another one, which is mostly because I love nature, so let me open it. **Michal Gula** (1:03:09): And it's **Florent Poux** (1:03:10): still driven by, you know, 3D Resona. so this is a cool one. You, you have trees. Yeah. It's on a slope, so it's not easy. And it's application that I still get asked a lot about forestry. And this is complex because you have all the trees very close to one another. So classical Euclidean-based clustering will fail here because everything is connected. So this is why, in this kind of approaches to get to this result, you need to be super smart about the way you, you, you approach that. and then as you can see, you are able to model all the trees one by one. And this is what is done at the moment. And you have all the forestry numbers, how many trees, how many trees per hectare, the biomass, everything is there. So this is a gold mine for anyone that works in forestry. and I love to show that because you can really see the various components and you can pick it. So if I go and go on the tree sides of things, or the stance or the model, maybe the export, you see, you can select each tree. and this is amazing because you have a very good delineation. It's very precise. So you were wondering about the precision of things. This is the kind of thing that you can have. And it speaks to the power of interpretation. Doing this job with old ways of doing things, so, wi- without having the smart point cloud infrastructure or so, it's very, very long. And today, th- this is all a matter of what I classify in, um- Smart, let's say, green, green tech, green tech and energy, right? And, I have other things with solar, but here it's very, very interesting. And you have... This is a lidar from above, but you have the same with drone from photos. The engine, you can use a train model or something else, and it's very interesting to always see the difference, and it talk to what you mentioned before. We compute score with geometry, and the train model will evaluate, or the train model will compute, and geometry w-will evaluate, or both together. I, **Michal Gula** (1:05:15): I forgot to ask this before, but, **Florent Poux** (1:05:19): is **Michal Gula** (1:05:19): it-- is your model working solely on geometry, or does it also need RGB colors? **Florent Poux** (1:05:26): it works with geometry only if you only have geometry, but it uses all available features. So if you have intensity, it will leverage that. If you have RGB, it will leverage that. If you have, 360 images, it will leverage that. If you have any things, you will be able to use it, yeah, in, in all the processes. This is multimodal, so essentially even in some aspect you will, you will get information, contextual information. This is what I have with the solar modeling. This is what we have with solar modeling. especially in France, depending on the region where you are, you will have different roof shapes or, or these kind of things. And having this contextual knowledge is fed to the structure so that the modeling approach is a bit different, depending on where you are. But you don't have for that to code anything. It's the, it's the model that has this capacity to go check on ontologies of existing, knowledge about regions and pick up what is interesting to be able to, let's say, adapt the way the, the, the roof will be constructed or, or, or the LoD3 model will be, all this kind of things. **Michal Gula** (1:06:40): there, there's this typical discussion in, in, reality capturing world, versus m-digital twins world. Where or what is a digital twin and what is not? And typical answer from digital twin guys is that, "Oh, your point clouds or 3D meshes, it's only a stupid geometry or stupid data. It's not a digital twin." Here we were able to see segmented data, connected to intelligent model or so-called intelligent, giving all the information. Let's say you have industrial plant imported in, in, in this solution and, giving you all the answers. Would you consider it a digital twin? **Florent Poux** (1:07:29): I'm careful here. yeah, it's, I mean, it's, for, for me it's taxonomy. If we want to be simple and not play on word, indeed, I agree that if you really want to have something that is reality in which you can operate, you don't need geometry only. You need all these aspects. and this aspect don't stop al-also at semantics. Having geometry semantics or class information is not enough. You need to have relationships. You need to have context. You need to have materials. you need to have physics. All of this to create something that you can actually extend on. And when I built the Smart Point Cloud, for example, I remember explicitly, at the end of the thesis, you know, you n- you, you say, "What do you believe will happen next?" So try to forecast a bit. And I remember that I said, "All right, so the next stage of evolution for the Smart Point Cloud is when we will have a natural language processing interface good enough so that we don't even need... Like, we can speak to it." And I'm super happy to see that it's, it was a valid, projection. But down the line, this structure, and it's not preaching for my church, it's, just saying that however you call it, you just need this component or this meta model that allows you to, to let the raw data, raw sensory input talk with whichever, input, question from the u- the end user, and that's it. And if you have that, you can call it a digital twin. I'm personally fine with it. so yeah. A-and this is another example, right? Where it, uh- **Michal Gula** (1:09:15): That's, that's safe and fair answer. okay. The, the very natural question from the audience, here will be: How can we try it? **Florent Poux** (1:09:29): Yeah. So, how can you try it? The, the Neurones 3D Science part of thing is accessible to people in the program, the 3D AR Architect program, where you have everything from 3D reconstruction to segmentation to Smart Point Cloud to scan to BIM, et cetera. **Michal Gula** (1:09:49): if you want- So what, what they, what will you learn in this program? **Florent Poux** (1:09:52): So they learn essentially to, create products or create, build research, that is very impactful to other. So you can come up without a defined plan. This is fine, because in the first session I give exactly the markets, where we are at, what is important, what is missing, and so on. So you can build your business from there. But m- a lot of the students in it, they are actually, engineers or they are founders of companies, and they, they really want to build, features without being tied to a specific software they use or they want to own it and be able to extend it, which in my opinion is the best possible scenario for them to be in, because then they're in control of costs, everything that rela- revolve around it. Um- And the pillars that are targets are, 3D reconstruction, baking understanding in your data, and automatic extraction of deliverables. Today, the way I teach is not the same as how I teached for years. That's also a reality. Before, I was going very deep in the thing, and today I make sure that the theory is enough to get quick into action. And because they have access to these tools, it's just a matter for them to use their favorite AI if they want to just prompt whatever they want to build on top, and they, and they will have it essentially. So then there is this duality where I actually at the moment teaching both the student and the agent of the student, which is a bit weird. Uh- So **Michal Gula** (1:11:31): basically, everything you were showing today, your students are, are able to learn in your program, right? **Florent Poux** (1:11:38): Yeah, exactly. Exactly. **Michal Gula** (1:11:40): is it, is it the program that you are teaching live or is it somewhere stored, people can re-watch it anytime they want? How does it work? **Florent Poux** (1:11:49): They are both. to protect at the moment because I, I push a lot, it's pre-recorded, so they can watch on their own time. But usually I push specific sessions as well, and we have a monthly session where we deep dive in a specific topic of someone or help or create a, a product altogether, essentially. There's a, there's a very cool thing that I love to do. and aside from that, if people want to, I don't know if MC is screen sharing. **Michal Gula** (1:12:17): Sure. They **Florent Poux** (1:12:17): want to pick up the, Rust version of it, there is this Neurones 3D engine, it's, it's free, where y-you, you can have it. You have everything about point cloud, Gaussian splat, rendering model, whatever. and you can use it directly so that you can check most of the features. I didn't push the last version, but, you already have a lot there. It's free. Uh- **Michal Gula** (1:12:39): Okay. So, so even people that are not in a program can, can try the- **Florent Poux** (1:12:42): Yeah, yeah, exactly. But, but this is only the part, it looks different. it's really when something is, is, sound and it's okay, it goes there. So for example, in AI analysis, you will have some holes, indoor power line, outdoor, whatever, and you will have the scan to plan, you will have some components, but, but not all the latest stuff that is on the science part. **Michal Gula** (1:13:06): Okay. But, within the program, you are not offering something like, like an engine where you build your own solutions. You are just teaching people how to use different, models and, and, how to basically vibe code their own solution, right? Like, m-my, like I'm aiming to, to one point, which Arek Szadkowski was actually Asking whether if you build something within your program, how will you keep it, like, updated? Because this, this is developing very fast. So in a half year there will be new things, new updates, and everything else. So how people joining your program will **Florent Poux** (1:13:48): be **Michal Gula** (1:13:49): able to keep up-updated with the, with the information? **Florent Poux** (1:13:54): I mean, I mean it's always the same. we tend to forget with hype that fundamentals are everything. So one thing that has not changed since, and, since it's there, so since five years, is there are core component like 3D deep learning, segmentation, spatial AI as well, so how to build agents and so on. All these components, they share methodological framework that is much more valuable, that's the actual implementation. So it's decoupled. So what may change the implementation, but this is fine because for me, just an update, there, there is, I don't know, a new library, a new way to do things, or something is not relevant anymore. It's very easy to update. but the methodological ground is something that I think is the most important in the course. and this is something that is like that. So **Michal Gula** (1:14:48): the, the, the so the logic and architecture of building the, the solution shouldn't change, only the tools you will need, to use, right? **Florent Poux** (1:14:59): but sometime the tools shape also the methodology somehow, so there is still this ambivalence to-... to figure out. But yeah, definitely like if we take a stupid example, w-we will deal with data always. So you need to know what a point cloud is, what are the, the parameters, what you need to take care of in 3D reconstruction, what kind of approach do you want? It's always a matter... For, for me, it's really about delegating. If you delegate to AI, it's best if you have a, a degree of understanding so that you can actually scope properly from the get-go. you, you, you can estimate the price of something, you can estimate the length it will take. And for having this, you need this high level view on the topics at hand. And if we talk about 3D and spatial AI and reconstruction and scan-to-BIM, yeah, th-this will not, specifically change. You, you need this kind of level of education. Now, implementing it, right? So do you, do you use, K-means over a DBSCAN over something else? This is something that may change. You, you, you may not need anymore to know the distinction in between. But the day, a system breaks, it will be super handy that you have been trained to understand how things are, are processed. again, this is something that I share with all the students in the academy is this, we like to be in control somehow of our own knowledge. Take the time, spend the time to learn things the right way, and it's okay to breathe a bit in this, area where there is a, a novelty every day. I think you feel much more in control when you feel yourself equipped at a very technical level on all the topic at hands. But again, I'm not a like... Like here, the, the idea was not to, to preach for, for getting and hopping on the program. It's really, it's really just more a matter of, showing what you can do today if you structure correctly your data. **Michal Gula** (1:17:07): Florent, thank you very much for this. I think this is a nice way how to wrap up this conversation because what you showed us, what's possible with our 3D data is, is really amazing. And as I'm always saying, AI is still, a toddler, so we can only hope and, and wait for what's going to happen. **Florent Poux** (1:17:27): Or, or you can build the next thing that is going to happen. So that's also the thing. In the academy, you are able to build your own agents, you are able to build your own AI, so that you don't rely on Claude, on ChatGPT or whatever because right now they are operating at a very, very low cost. They are making absolutely no margin. They are losing money. But the day they stop doing that, the cost will not be the same token-wise. So it's very important to know that so that you're prepared for, for that shift, and you have a backup plan or you already put in place solution that you own, including your own AI. And today you can do that, so yeah. **Michal Gula** (1:18:08): guys, you heard it. if you wanna learn how to build this, this, segmenting tool and to make your data, actually smart, not stupid, as they are today, get in touch with Florent. go and visit his website. Join his program. you could hear it for last 80 minutes that he's a very nice guy to talk to. So go ahead and, and learn something new. Thank you, Florent, for doing this. **Florent Poux** (1:18:36): Thanks a lot, Michal. Thanks everyone. Bye **Michal Gula** (1:18:39): guys. Cheers.