surrey: a PointNet++ style lidar segmentation model trained on DALES-2

Weights of the models in the project repository, github.com/markolivaic/surrey. The README there gives every figure, each with its null and its seed range; this card gives none, so it cannot drift from them.

Files

File	What it is
m0_seedA.pt	The published model, M0 seed A: the one the page shows
m0_seedB.pt, m0_seedC.pt	The same recipe with two other seeds, the range behind every figure
m0_big_seedA.pt	The capacity control, a wider network with the same recipe; not adopted

Each file holds the network's tensors and a few plain values (channels, width, recipe, seed, step and the first 16 hex digits of the source training checkpoint's SHA-256). It loads with torch.load(..., weights_only=True), which is what surrey.inference.load_checkpoint does by default.

Input and output

• Input: points in 40 m blocks, three values each: x and y centred on the block, z above the block's floor, in metres. No intensity, no return numbers, no colour.
• Output: fifteen DALES-2 classes per point. The class names and their order are in backend/surrey/data.py of the repository.
• Inference averages the softmax over several seeded passes, each drawing its own sample centroids and its own dealing of points into blocks (backend/surrey/inference.py).

Training data and what it does not know

Trained on DALES-2 (Bendjilali and others, 2026), aerial lidar flown over Surrey, British Columbia, Canada. The labels have no class for water or bridges, so the model has no way to call either.

Run on USGS 3DEP lidar over San Francisco, the model's classes are predictions, not measurements, and not part of any USGS deliverable. How far they hold there, and where they do not, is the subject of the repository.

Licence

CC BY-NC 4.0. Every model here learned from the DALES-2 labels, which revise the DALES annotations, whose terms exclude commercial use; the restriction is passed on. LICENSE_MODEL.md in the repository sets out why and what to credit.