3D Semantic Segmentation on ScanNet (full)
60Training Latency (ms)MinkUNet
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| MinkUNet#Params=39.2M, GPU=RTX 4090, AMP=Enabled (Training) / Disabled (Inference)2025.12 | 60 | — | 21 | — | |
| LitePT-S#Params=12.7M, GPU=RTX 4090, AMP=Enabled (Training) / Disabled (Inference), Decoder=Standard (Linear projection)2025.12 | 72 | — | 21 | — | |
| LitePT-S*#Params=16.0M, GPU=RTX 4090, AMP=Enabled (Training) / Disabled (Inference), Decoder=Heavier (includes attention or convolutional blocks)2025.12 | 81 | — | 26 | — | |
| LitePT-B#Params=45.1M, GPU=RTX 4090, AMP=Enabled (Training) / Disabled (Inference)2025.12 | 93 | — | 33 | — | |
| LitePT-L#Params=85.9M, GPU=RTX 4090, AMP=Enabled (Training) / Disabled (Inference)2025.12 | 97 | — | 41 | — | |
| PTv3#Params=46.1M, GPU=RTX 4090, AMP=Enabled (Training) / Disabled (Inference)2025.12 | 110 | — | 51 | — | |
| PTv2#Params=12.8M, GPU=RTX 4090, AMP=Enabled (Training) / Disabled (Inference)2025.12 | 188 | — | 151 | — |