392×
less memory than a uniform 5 cm grid
94.9%
point accuracy vs ground truth
100%
points binned, zero data loss
45 ms
grid stage, inside a 10 Hz budget
Driving sequence
A real KITTI drive played back frame by frame. The map is fixed in world coordinates and the vehicle moves through it, with moving objects marked apart from parked ones.
Open the animation →Single scan report
One scan in full detail: the grid in four colour modes, surface defects, resolution rings, traversability and the driving advice that comes out of them.
Open the report →What the system does
- Foveated 2.5D grid — 5 cm cells within 10 m, widening to 50 cm at 100 m, because 70% of a scan's points fall in the near field and a uniform grid spends the same memory on empty distance.
- Semantic segmentation — SalsaNext over 20 SemanticKITTI classes, collapsed into drivable terrain, static obstacles and dynamic objects.
- Overhangs kept separate — a branch over a road no longer marks the road blocked; the clearance between them is reported in metres.
- Surface defects located — depressions and raised patches found by comparing each cell against its neighbours, not just flagged as rough.
- Reachable corridor — where the vehicle can actually get to, which is not everything labelled "road".
- Moving vs parked — a single scan cannot tell them apart; tracking across scans can.
Honest notes
- Every figure here is measured on real KITTI data, not estimated. Memory is compared against a uniform 5 cm 2.5D grid over the same 100 m radius.
- The 392× splits into 26× from foveation — the variable cell size, which is the actual contribution — and 15× from storing only occupied cells.
- Accuracy is 94.9% at the point level and 70.0% mIoU, scored against SemanticKITTI labels over 50 scans. mIoU is the harder number and the honest headline.
-
These pages are pre-rendered. The full pipeline, including uploading
your own scan, runs from the repository with
python server.py.