jmgoo1118/mls-cross-sensor-segmentation

Cross-Sensor MLS Semantic Segmentation: trained weights

Weights for the paper Cross-Sensor Generalization of MLS Point-Cloud Semantic Segmentation. Code, configs, and usage: https://github.com/junegoo94/mls-cross-sensor-segmentation

All models are Point Transformer V3 (Pointcept) trained on coordinates only (in_channels=3) and predict the unified 11-class label space. Each file is a weights-only checkpoint ({"state_dict": ...}, optimizer removed, module. prefix stripped, about 185 MB) that loads with the matching config and strict=True.

File	Trained on	Config	Used in
single_toronto3d_coord.pth	Toronto3D	semseg-pt-v3-coord.py	in-domain, 3x3 transfer
single_parislille3d_coord.pth	Paris-Lille-3D	semseg-pt-v3-npm3d-coord.py	in-domain, 3x3 transfer
single_whuurban3d_coord.pth	WHU-Urban3D	semseg-pt-v3-whu-coord.py	in-domain, 3x3 transfer
loso_heldout_toronto3d_recipeB.pth	Paris-Lille-3D + WHU-Urban3D, Recipe B	semseg-pt-v3-loso1b-AB-coord.py	LOSO, held-out Toronto3D
loso_heldout_parislille3d_recipeB.pth	Toronto3D + WHU-Urban3D, Recipe B	semseg-pt-v3-loso2b-AB-coord.py	LOSO, held-out Paris-Lille-3D
loso_heldout_whuurban3d_recipeB.pth	Toronto3D + Paris-Lille-3D, Recipe B	semseg-pt-v3-loso3b-AB-coord.py	LOSO, held-out WHU-Urban3D
selftrain_heldout_toronto3d.pth	as above + pseudo-labelled Toronto3D	semseg-pt-v3-uda-toronto-coord.py	self-training reference
selftrain_heldout_parislille3d.pth	as above + pseudo-labelled Paris-Lille-3D	semseg-pt-v3-uda-npm3d-coord.py	self-training reference
selftrain_heldout_whuurban3d.pth	as above + pseudo-labelled WHU-Urban3D	semseg-pt-v3-uda-whu-coord.py	self-training reference
deploy_threesource_recipeB.pth	all three sources, Recipe B	semseg-pt-v3-all3-AB-coord.py	deployment on a new sensor

The model self-trained on the proprietary Ordnance Survey point cloud is not released.

Unified classes (index order): ground, building, natural, pole, car, fence/barrier, road marking, utility line, bollard, trash can, pedestrian.