3D Intelligence Report – July 28, 2026
Theme of the day: 3D data outgrew the machines that hold it, and the answer showing up everywhere is the same one. Cut the problem into pieces instead of buying bigger hardware. One paper partitions an outdoor drive rather than holding the scene whole. One tool rewrites its core so memory is bounded by chunk size instead of scene size. One market move ties positioning hardware to point cloud processing, a bet that the classified result, not the instrument, is where the value settles.
Every link below was fetched and verified on July 28, 2026, the day this report went out.
GLAM-SLAM: Real-time Gaussian Large-scale Mapping via Flow Densification and Spatial Decomposition
Panagiotis Mermigkas, Argyris Manetas, Petros Maragos
about 15% better reconstruction quality than the second-best system, at real time, on kilometre-scale outdoor drives
Gaussian splatting SLAM has mostly been an indoor, room-scale story because a room fits in memory. GLAM-SLAM decouples the problem: a feature-based SLAM frontend handles localization, sparse anchor grids plus MLP-initialized scene partitions handle the map, and epipolar-constrained flow densification gives the Gaussians geometry to sit on where views are thin. Evaluated on KITTI Odometry, Oxford RobotCar and Malaga long sequences.
The score is not what interests me here, the decomposition is. Splatting SLAM keeps hitting the same wall: it was designed for scenes you can hold whole, and an outdoor drive is not one of those. Cutting the map into partitions with their own local Gaussians is the move that makes the method survive contact with a real vehicle route. I want to see the code before I trust the numbers, and today there isn't any, just a README and a template project page. But the architecture is the right answer to the right problem.
Principal Engineer, 3D Reconstruction
$230,000-$310,000 base (est.)
This is not a role that ships a product, and that's exactly what makes it interesting. Wayve wants an offline reconstruction system that turns raw vehicle sensor data into trustworthy 3D geometry and ego motion, so that validation, autolabelling and model training all have something reliable to sit on. Look at what they actually ask for: standards for reconstruction quality, failure analysis, uncertainty measurement, credibility arguments for reconstructed data. Not make it look right, prove how wrong it might be. That's a surveyor's discipline wearing a robotics job title, and it's a reminder that the skills sitting in geodesy and photogrammetry departments are increasingly what the AV industry is short of.
Four funding doors, four regions
SplatTransform v3.x line (v3.0.0 core rewrite, v3.1.0 streaming, v3.1.7 current)
memory bounded by chunk size, not scene size
v3.0.0 replaced the core with a streaming ChunkSource: memory is bounded by chunk size, not scene size, so conversion no longer requires the scene to fit in RAM. CLI flag names were rationalized in the same release, so v2-era scripts need reviewing. v3.1.0 added streamed SOG loading and reworked voxel processing for large scenes. v3.1.7 loads pre-v1 lod-meta.json and reports input format plus extra columns in –info / –stats.
This is the least glamorous kind of release and one of the most useful. A converter that reads the whole scene into memory quietly makes your RAM the legal limit on the size of capture you're allowed to own. Streaming moves that ceiling from what fits to how long you're willing to wait, which is a much better constraint to be under. It builds directly on the v2.6.0 entry from June: same tool, but the June news was a new format to read, and this is the machinery underneath being rebuilt. Check your flags before you rerun old scripts.
Strategic partnership joining Topcon's positioning hardware and new geomatics business unit to GreenValley's AI point cloud processing stack, across handheld, aerial and mobile LiDAR
Topcon has stood up a dedicated geomatics business unit with investment in AI, cloud connectivity and data unification. GreenValley brings automated feature extraction and classification. The named targets are surveying, mapping, construction and forestry, with robotic systems, autonomous monitoring and real-time cloud processing on the roadmap. Ron Oberlander (Topcon) framed the goal as helping users collect, process and apply data more effectively, including in remote and challenging environments.
Watch the direction rather than the wording. A positioning-hardware company is buying its way into the part of the workflow that happens after the scan, because that is where the margin went. The instrument is becoming the commodity end of the deal and the classified, queryable, cloud-hosted result is becoming the product. If you sell processing services on point clouds, this tells you which part of your work the hardware vendors plan to absorb, and it is worth deciding now whether you compete there or move up into the interpretation nobody has automated yet.
Get the next report in your inbox
Five verified finds, my take on each, one short email a day.
Five verified finds with my take, one short email a day.
