Apparently, the 3rd Wednesday of the 3rd week of the 3rd month is 𝟑𝐃 𝐃𝐚𝐲. Had you heard about it? 👀 Seems like we're a bit late - but it's never too late to talk about the most exciting format in 3D right now: 𝐆𝐚𝐮𝐬𝐬𝐢𝐚𝐧 𝐒𝐩𝐥𝐚𝐭𝐭𝐢𝐧𝐠.  Photorealistic, real-time scenes you can navigate freely from any angle. And what makes it interesting for us: it works beautifully with 360° images - a format that serious mapping professionals weren't exactly rushing to adopt. Until now. During our recent webinar, Tomas Barnas, Michal Gula, and Jeffrey Martin shared how their own perspective on 360° imaging shifted completely once Gaussian Splatting entered the picture. Have you ever felt that way about a technology? Dismissed it, only to watch it completely change the game? 📹 Find the full webinar on our YouTube channel: https://lnkd.in/dcuNbZgd Curious about what Gaussian Splatting is actually capable of and where it's heading next? We sat down with our Lead Computer Vision Engineer for a full Q&A on the topic. Check out the link below 👇 #3DDay #MobileMapping #Geospatial #RealityCapture #Photogrammetry

Why your perception algorithms may fail or perform poorly in lunar environments: In the clip, we see two versions of DepthSplat, a "feed‑forward" 3D Gaussian splatting architecture that relies on monocular depth estimation (MDE) using Depth Anything V2. The pre‑trained model was trained on a diverse collection of indoor–outdoor datasets, including TartanAir (a synthetic dataset with indoor and outdoor robot environments), RealEstate10K (real‑world mostly‑indoor home‑tour videos used for Gaussian splatting and unsupervised depth pretraining), ScanNet (real indoor RGB‑D scans of apartments and offices), DL3DV‑10K (a large real‑world dataset with indoor and outdoor scenes), KITTI (real outdoor driving scenes used for zero‑shot depth evaluation), and VKITTI2 (a synthetic clone of KITTI used for depth fine‑tuning). Despite this diversity, the pre‑trained model performs poorly in the left clip—there are visible distortions on the pillars and especially around the glass surfaces. On the right, the same architecture is shown again, but this time trained from scratch using our lunar‑testbed‑specific data. The resulting inference is significantly better: no major distortions, no ballooning splats around the testbed’s glass elements, and overall more stable geometry. I used only four views and low‑resolution images to validate the post‑training performance, so the photorealism is not perfect, but the improvement is unmistakable. This provides a vivid example of why extreme environments—such as lunar terrain, where the lack of atmosphere allows a persistent, asymmetric dust cloud to form—can easily break even the most robust perception algorithms. This issue is not limited to MDE (which is actually a strong option here) but also affects common IR‑based LiDARs and depth cameras. We have repeatedly observed IR rays scattering off regolith dust and producing stochastic noise patterns that do not appear in everyday outdoor environments or in the datasets typically used for training. This is why a lunar‑analog site or a controlled artificial simulation testbed is essential when developing perception algorithms intended for the Moon. #robotics #gaussiansplats #3dreconstruction

If 𝐬𝐜𝐞𝐧𝐞 𝐫𝐞𝐜𝐨𝐧𝐬𝐭𝐫𝐮𝐜𝐭𝐢𝐨𝐧 starts anywhere, it starts with points. → Before meshes. → Before semantics. → Before visual realism. Just points in space. ⸻ 1️⃣ 𝐅𝐞𝐚𝐭𝐮𝐫𝐞 𝐏𝐨𝐢𝐧𝐭𝐬 (𝐒𝐩𝐚𝐫𝐬𝐞 𝐆𝐞𝐨𝐦𝐞𝐭𝐫𝐲) Many reconstruction systems begin by detecting and matching feature points across frames. 𝘾𝙤𝙧𝙣𝙚𝙧𝙨. 𝙀𝙙𝙜𝙚𝙨. 𝘿𝙞𝙨𝙩𝙞𝙣𝙘𝙩 𝙩𝙚𝙭𝙩𝙪𝙧𝙚𝙨. These are: • Stable under viewpoint changes • Easy to match across images • Useful for estimating camera pose This gives you a sparse point cloud. Enough to understand the structure. Not enough to represent a room. ⸻ 2️⃣ 𝐃𝐞𝐩𝐭𝐡-𝐁𝐚𝐬𝐞𝐝 𝐏𝐨𝐢𝐧𝐭 𝐂𝐥𝐨��𝐝𝐬 (𝐃𝐞𝐧𝐬𝐞 𝐆𝐞𝐨𝐦𝐞𝐭𝐫𝐲) To go from sparse to dense, you need depth. On mobile, that can come from: • LiDAR (direct depth sensing) • Structured light (older devices) • Stereo / multi-view geometry • Or learned depth models LiDAR gives you immediate depth measurements. → It’s fast. → It’s practical. → But it’s noisy and resolution-limited. The result? 𝗔 𝗱𝗲𝗻𝘀𝗲 𝗰𝗹𝗼𝘂𝗱 𝗼𝗳 𝗽𝗼𝗶𝗻𝘁𝘀 𝗶𝗻 𝟯𝗗 𝘀𝗽𝗮𝗰𝗲. In the attached video, we’re showing a raw point cloud of our own office, captured entirely with a phone → Each point = (x, y, z) → Sometimes with color. → Often with noise. 𝙄𝙩’𝙨 𝙖 𝙢𝙚𝙖𝙨𝙪𝙧𝙚𝙢𝙚𝙣𝙩 𝙤𝙛 𝙨𝙥𝙖𝙘𝙚 - 𝙣𝙤𝙩 𝙖𝙣 𝙞𝙣𝙩𝙚𝙧𝙥𝙧𝙚𝙩𝙖𝙩𝙞𝙤𝙣. ⸻ 3️⃣ Learned Depth (Depth Anything & SOTA Models) In recent years, state-of-the-art depth models like Depth Anything have made monocular depth surprisingly strong. From a single RGB image, you can infer relative depth with impressive quality. But there are tradeoffs: • Scale ambiguity • Temporal inconsistency • Sensitivity to lighting and domain shift On mobile, this introduces additional constraints: → Model size → Inference latency → Battery consumption Depth from sensors and depth from models both produce point clouds. But they differ in: • Accuracy • Stability • Computational cost ⸻ The Core Limitation Even with perfect depth: A dense point cloud is still just a collection of samples. → No surfaces. → No topology. → No semantic meaning. 𝙄𝙩’𝙨 𝙧𝙚𝙖𝙡𝙞𝙩𝙮 𝙢𝙚𝙖𝙨𝙪𝙧𝙚𝙙 - 𝙣𝙤𝙩 𝙧𝙚𝙖𝙡𝙞𝙩𝙮 𝙪𝙣𝙙𝙚𝙧𝙨𝙩𝙤𝙤𝙙. And that’s why every reconstruction pipeline has to move beyond it. Next: how we impose structure on chaos - and turn measurements into meaning. 🔥 #MobileVision #SpatialComputing #ARDevelopment #MobileEngineering #MobileInnovation

Beyond bare earth: Could laser return intensities provide another layer of data for Search and Recovery? First off, it is a very good thing I am friends with my neighbors. If not, the local PD might have paid me a visit after seeing me in the front yard setting out plastic skeletons, tarps, carpets, commercial trash bags, and a shovel! All jokes aside, this setup served a highly specific purpose. I spent some time today practicing with point cloud data manipulation, specifically analyzing the similarities and differences in laser return intensities across various materials. I am continually exploring how these concepts might be applied to missing person recovery and cold cases. Consider how we currently utilize LiDAR in SAR: Topography: We know LiDAR can provide an excellent depiction of the bare earth, helping investigators identify unnatural mounds or depressions hidden under the canopy. Material Signatures: What if we could also manipulate the data to isolate specific laser intensities when a certain material is known to be involved? In this particular scan, I was focusing on the return signature of a commercial trash bag (upcoming cold case scan). Placing a known reference material in the scan area is certainly helpful to establish a baseline (but still must consider the many factors that may affect the laser pulses: angle, wet, dry, weather, etc.). Operating with a "trust but verify" mindset means acknowledging where the technology currently stands. Analyzing intensities might not provide an exact "find it" moment in the same way thermal anomalies or color pixel recognition software might. However, as hardware, software, and our own ongoing education and practice continue to evolve, understanding material reflectivity could become yet another valuable tool for investigators. We have to keep testing the limits of our equipment before we ever reach a critical search area. 🔗 Consider how advanced reality capture and data manipulation might support your next complex investigation at www.scenephoto360.com. #LiDAR #SearchAndRecovery #ColdCase #ForensicScience #PointClouds #DataManipulation #ScenePhoto360 #GIS #PublicSafety #TrustButVerify

We’re excited to introduce Get3D.AI’s City-Level Solution. Built for large-scale city reconstruction, this solution is designed to help teams manage massive datasets and deliver high-quality 3D results more efficiently. From multi-source data processing to reconstruction and final delivery, #Get3DAI provides an end-to-end workflow for complex city-scale projects—making large-scale production more streamlined, stable, and practical. Supporting aerial imagery, close-range capture, and point clouds, our City-Level Solution enables more complete 3D outcomes for city mapping, digital twin development, infrastructure management, and large-area visualization. 🔗：https://lnkd.in/g3iJ9hv8 https://www.get3d.ai/ Explore our new solution page and discover what’s possible with scalable city-level 3D reconstruction. #Get3DAI #CityLevelSolution #3DReconstruction #DigitalTwin #Geospatial #Photogrammetry #Surveying #Mapping #PointCloud #SmartCity #workflow #LargeScale

We are moving past the era of just rendering static scenes. The next frontier in experiential design and spatial storytelling is mathematically understanding the physical world in real time. Lately, I’ve been deeply obsessed with real-time data overlays and spatial analysis. Think about a tree blowing in the wind. Tracking and labeling every single leaf isn't just an aesthetic flex; it’s a massive computational stress test. It bridges raw computer vision with visual art, turning a physical object into a fluid dynamics data matrix of X, Y, and Z coordinates. Watching how sonar maps out hidden topographies sparked a realization: the true impact lies in visualizing the unthought-of. Whether it's through dense optical flow, Time-of-Flight sensors, or live LiDAR meshing, the ability to convert invisible physical systems into stunning, real-time data overlays is where the bleeding edge of our industry is heading. I want to see the things people rarely think about capturing. I want to connect with the people pushing these boundaries. If you are an engineer, visual artist, or creative technologist building in this space,what is the most complex real-time spatial data you are tracking right now? Let’s talk in the comments or dm me.

🤖 Post #617: arXiv:2603.03265 **DuoMo: Dual Motion Diffusion for World-Space Human Reconstruction** Recovering how a person moves through the world — not just on screen — requires two models thinking in two spaces. Key contributions: • Camera-space diffusion model generates initial motion estimates from video input • World-space diffusion model lifts and refines estimates for global consistency • Works directly on mesh vertex positions, no parametric body model required • 30% world-space error reduction on RICH, 16% on EMDB (CVPR 2026) #MachineLearning #HumanMotion #ComputerVision #AIResearch

🇩🇴 Live from the Dominican Republic: A New Era of Surveying with SLAM RTK! Thrilled to connect with the surveying community in the Dominican Republic! During our recent user meeting, we demonstrated the incredible capabilities of the KOLIDA ME Plus handheld SLAM RTK. 🚀 Watch how it transforms complex physical spaces into precise digital point clouds in seconds! From 3D modeling to structural monitoring, the portability and efficiency of the ME Plus left local experts truly impressed. Surveying is no longer about tedious setups—it's now a "what you see is what you get" walking experience. Check out the video below to feel the energy and see the tech in action! 👇 #KOLIDA #SLAMRTK #Surveying #Geospatial #DominicanRepublic #Innovation #SurveyLife #DigitalTwin

Yesterday, we talked about where 𝐬𝐜𝐞𝐧𝐞 𝐫𝐞𝐜𝐨𝐧𝐬𝐭𝐫𝐮𝐜𝐭𝐢𝐨𝐧 begins: points in space. Sparse features, LiDAR depth, and learned depth models all lead to the same thing - a dense point cloud. But a point cloud is still 𝗷𝘂𝘀𝘁 𝗮 𝗺𝗲𝗮𝘀𝘂𝗿𝗲𝗺𝗲𝗻𝘁 𝗼𝗳 𝘀𝗽𝗮𝗰𝗲. Today, we move to the next step → 𝘁𝘂𝗿𝗻𝗶𝗻𝗴 𝘁𝗵𝗼𝘀𝗲 𝗽𝗼𝗶𝗻𝘁𝘀 𝗶𝗻𝘁𝗼 𝘀𝘂𝗿𝗳𝗮𝗰𝗲𝘀. ⸻ Once you have a dense point cloud, the next step is structure. A mesh connects points into 𝘁𝗿𝗶𝗮𝗻𝗴𝗹𝗲𝘀 and 𝗰𝗼𝗻𝘁𝗶𝗻𝘂𝗼𝘂𝘀 𝘀𝘂𝗿𝗳𝗮𝗰𝗲𝘀. Now you don’t just measure space - 𝘆𝗼𝘂 𝗱𝗲𝗳𝗶𝗻𝗲 𝗶𝘁. The attached video shows a textured mesh of our office, reconstructed entirely from a phone capture. ⸻ 1️⃣ 𝐇𝐨𝐰 𝐦𝐞𝐬𝐡𝐞𝐬 𝐚𝐫𝐞 𝐛𝐮𝐢𝐥𝐭 Common approaches include: → Poisson surface reconstruction → Delaunay triangulation → TSDF / volumetric fusion All of them try to infer continuous surfaces from discrete samples. 𝘽𝙪𝙩 𝙗𝙚𝙛𝙤𝙧𝙚 𝙢𝙚𝙨𝙝𝙞𝙣𝙜, 𝙩𝙝𝙚 𝙥𝙤𝙞𝙣𝙩 𝙘𝙡𝙤𝙪𝙙 𝙪𝙨𝙪𝙖𝙡𝙡𝙮 𝙣𝙚𝙚𝙙𝙨 𝙬𝙤𝙧𝙠. Typical preprocessing steps: → Outlier removal → Depth smoothing/averaging → Normal estimation → Frame alignment refinement If your point cloud is noisy or misaligned, the mesh will amplify those artifacts. 𝙂𝙖𝙧𝙗𝙖𝙜𝙚 𝙞𝙣 → 𝙨𝙩𝙧𝙪𝙘𝙩𝙪𝙧𝙚𝙙 𝙜𝙖𝙧𝙗𝙖𝙜𝙚 𝙤𝙪𝙩. ⸻ 2️⃣ 𝐀𝐑𝐊𝐢𝐭 𝐒𝐜𝐞𝐧𝐞 𝐑𝐞𝐜𝐨𝐧𝐬𝐭𝐫𝐮𝐜𝐭𝐢𝐨𝐧 On LiDAR-enabled iOS devices, ARKit provides real-time mesh reconstruction. It fuses depth data into a triangle mesh as you scan. It’s practical and fast - but: → Limited to LiDAR devices → Resolution is constrained → Memory usage grows quickly → Geometry is optimized for interaction, not precision It’s a usable mesh, 𝗻𝗼𝘁 𝗮 𝗽𝗲𝗿𝗳𝗲𝗰𝘁 𝗼𝗻𝗲. ⸻ 3️⃣ 𝐓𝐡𝐞 𝐌𝐨𝐛𝐢𝐥𝐞 𝐂𝐨𝐧𝐬𝐭𝐫𝐚𝐢𝐧𝐭 On mobile, meshing competes with: → GPU memory → Battery → Real-time performance Every triangle has a cost. That’s why the quality of your captured depth - and how you preprocess it - 𝗺𝗮𝘁𝘁𝗲𝗿𝘀 𝗺𝗼𝗿𝗲 𝘁𝗵𝗮𝗻 𝘁𝗵𝗲 𝗺𝗲𝘀𝗵𝗶𝗻𝗴 𝗮𝗹𝗴𝗼𝗿𝗶𝘁𝗵𝗺 𝗶𝘁𝘀𝗲𝗹𝗳. ⸻ But even with a clean mesh, there’s another question: 𝘿𝙤 𝙬𝙚 𝙬𝙖𝙣𝙩 𝙛𝙖𝙞𝙩𝙝𝙛𝙪𝙡 𝙜𝙚𝙤𝙢𝙚𝙩𝙧𝙮 - 𝙤𝙧 𝙖𝙣 𝙞𝙣𝙩𝙚𝙧𝙥𝙧𝙚𝙩𝙚𝙙 𝙧𝙤𝙤𝙢? In the next post, we’ll look at how systems like RoomPlan move from raw geometry to semantic structure. #MeshReconstruction #SceneReconstruction #3DReconstruction #PointCloud #SpatialComputing #ARKit #LiDAR #MobileAR #iOSDevelopment #RoomPlan

The End of 2D Broadcasting: Volumetric Optics 🏀 Current sports broadcasting relies on 2D pixel arrays, giving us a "flat" representation of a three-dimensional world. The first pillar of what I call the "Expansion" changes this completely by shifting to Volumetric Optics. In this new model, the camera evolves from a passive lens into an active LIDAR and spatial computing sensor array. It is designed to capture not just light, but vectors, velocity, and mass. Watch the video one here. When that basketball player leaps, the system isn't recording a traditional video of the jump; it records the exact physics of the movement: Vector v: The direction and speed. Mass m: The force exerted on the floor. Spatial Coordinates (x, y, z): The exact volumetric displacement of air. This data isn't meant for a standard screen. It is meant for a Physical Physics Engine. Part 2 is coming soon, where we explore the Faster-Than-Light (FTL) transmission of this physics blueprint straight to a robotic receiver. #SpatialComputing #FutureOfMedia #VolumetricOptics #SportsTech #ArtificialIntelligence #PhysicsEngine #NextGenMedia