3D Semantic Segmentation on ScanNet200 v1 (val)
51.51mIoU (Head)Ours
Evaluation results
| Method | Links | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| OursBackbone=Sparse 3D U-Net (80M MinkowskiNet), Pre-training=Language-grounded 3D pre-training2022.04 | 51.51 | 22.68 | 12.41 | 28.87 | 72.72 | 66.69 | 58.3 | 65.9 | 62.5 | 29.09 | 26.61 | 39.4 | |
| Ours (CLIP only)Backbone=Sparse 3D U-Net (80M MinkowskiNet), Pre-training=CLIP only2022.04 | 50.39 | 22.84 | 10.1 | 27.73 | 71.64 | 69.72 | 44.47 | 61.94 | 62.2 | 29.37 | 17.35 | 36.16 | |
| CSCBackbone=Sparse 3D U-Net (80M MinkowskiNet), Pre-training=Contrastive Scene Contexts2022.04 | 49.43 | 19.52 | 10.28 | 26.41 | 70 | 67.75 | 40.78 | 59.51 | 61.01 | 25.75 | 17.62 | 34.79 | |
| SupConBackbone=Sparse 3D U-Net (80M MinkowskiNet), Pre-training=Supervised Contrastive Learning2022.04 | 48.55 | 19.17 | 10.34 | 26.02 | 69.52 | 65.42 | 40.62 | 58.52 | 60.27 | 26.28 | 19.14 | 35.23 | |
| Ins. samp.Backbone=Sparse 3D U-Net (80M MinkowskiNet), Technique=Instance sampling2022.04 | 48.46 | 18.97 | 9.22 | 25.49 | 70.04 | 62.98 | 49.41 | 60.81 | 59.64 | 24.66 | 19.25 | 34.52 | |
| ScratchBackbone=Sparse 3D U-Net (80M MinkowskiNet), Pre-training=None2022.04 | 48.29 | 19.08 | 7.86 | 25.02 | 68.81 | 66.29 | 39.88 | 58.32 | 60.45 | 25.5 | 15.06 | 33.67 | |
| C-FocalBackbone=Sparse 3D U-Net (80M MinkowskiNet), Technique=Class-balanced Focal Loss2022.04 | 48.1 | 20.28 | 9.38 | 25.86 | 68.1 | 65.64 | 47.43 | 60.39 | 60.08 | 26.28 | 19.14 | 35.48 |