Soft Query Knowledge Distillation
This model provides the main distillation experiment results on dataset ScannetV2 and S3DIS for the 3D instance segmentation task in the model zoo, along with the inference code for the Mask3D model and the SPFormer3D method. The table below summarizes the main distillation results from our experiments. You can click on ckpt to download the corresponding Baseline model and the distilled model for each experimental group.
| ID | Baseline mAP / mAP50 / mAP25 | Comp.Ratio | Flops | SQKD mAP | SQKD mAP50 | SQKD mAP25 | download |
|---|---|---|---|---|---|---|---|
| 1 | 24.6 / 44.5 / 65.4 | *4.17 | 83% | 30.6 | 52.0 | 70.6 | ckpt1 |
| 2 | 37.7 / 63.0 / 79.0 | *4.17 | 83% | 41.0 | 66.3 | 81.2 | ckpt2 |
| 3 | 29.7 / 48.8 / 66.5 | *2.32 | 95% | 32.2 | 53.8 | 71.1 | ckpt3 |
| 4 | 43.0 / 68.2 / 81.9 | *2.32 | 95% | 46.0 | 71.4 | 84.0 | ckpt4 |
| 5 | 42.2 / 68.9 / 83.1 | *2.32 | 76% | 47.2 | 73.7 | 86.4 | ckpt5 |
| 6 | 42.4 / 68.0 / 82.3 | *2.32 | 57% | 47.6 | 72.6 | 84.1 | ckpt6 |
| 7 | 36.8 / 54.0 / 66.6 | *2.44 | 91% | 41.1 | 58.6 | 71.1 | ckpt7 |
| 8 | 38.3 / 56.3 / 68.4 | *2.44 | 91% | 41.7 | 59.6 | 70.8 | ckpt8 |
| 9 | 37.2 / 54.9 / 66.6 | *2.44 | 71% | 40.1 | 57.6 | 68.0 | ckpt9 |
| 10 | 25.1 / 40.6 / 53.1 | *3.87 | 80% | 30.3 | 45.8 | 57.7 | ckpt10 |
| 11 | 19.9 / 27.1 / 39.4 | *4.17 | 83% | 23.2 | 30.8 | 42.0 | ckpt11 |
| 12 | 24.5 / 37.8 / 52.4 | *4.17 | 83% | 30.8 | 43.3 | 56.9 | ckpt12 |
| 13 | 22.0 / 30.8 / 45.4 | *2.32 | 95% | 27.4 | 36.4 | 47.6 | ckpt13 |
| 14 | 33.3 / 44.3 / 55.3 | *2.32 | 95% | 38.9 | 49.9 | 59.4 | ckpt14 |
In the table, experimental groups 1–6, 7–10, and 11–14 use different teacher models, respectively. We also fixed some bugs in the label category code during data preprocessing and provide the improved versions of the Mask3D and SPFormer models.
For model environment setup and dataset preprocessing details, please refer to the original paper Mask3D_git and SpFormer_git.