Performance
June 24, 2022 ยท View on GitHub
SSOD
Fair Comparison
For fair comparison, we reproduce some SSOD methods with the same setting (e.g., augmentation, training iterations, batch-size, and so on.).
- Soft-Teacher: without box-jitter (we will add box-jitter in next version)
- PASCAL-VOC: use 2-GPU
COCO-standard
| method | batch-size | iterations | 1% | 5% | 10% |
|---|---|---|---|---|---|
| STAC (thr=0.9, CE) | 32 labeled + 32 unlabeled | 40K | 16.1 | 24.0 | 28.1 |
| Unbiased-Teacher (thr=0.7, FL) | 32 labeled + 32 unlabeled | 40K | 22.0 | 28.6 | 32.1 |
| Soft-Teacher (thr=0.9, CE) | 32 labeled + 32 unlabeled | 40K | 22.1 | 29.0 | 32.7 |
| LabelMatch (original code) | 32 labeled + 32 unlabeled | 40K | 24.6 | 31.5 | 34.6 |
| LabelMatch (here) | 32 labeled + 32 unlabeled | 40K | 24.6 | 31.6 | 34.4 |
NOTE: we use the ablation training setting here. (different training setting can be found in supplementary materials.)
PASCAL-VOC
| method | batch-size | iterations | AP50:95 | AP50 |
|---|---|---|---|---|
| STAC (thr=0.9, CE) | 8 labeled +8 unlabeled | 80K | 46.5 | 78.6 |
| Unbiased-Teacher (thr=0.7, FL) | 8 labeled +8 unlabeled | 80K | 53.3 | 84.2 |
| Soft-Teacher (thr=0.9, CE) | 8 labeled +8 unlabeled | 80K | 52.8 | 84.3 |
| LabelMatch | 8 labeled +8 unlabeled | 80K | 54.7 | 84.8 |