Object Detection on OpenImages v4 (val)
69.9mAP50 (Box)Dataset-specific oracles
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Dataset-specific oraclesSupervision source=Dataset-specific box labels, Evaluation protocol=Supervised2022.01 | 69.9 | 81.8 | — | |
| DeticBackbone=Swin-B, Supervision source=IN-21k (ImageNet-21k) image-level, Evaluation protocol=Zero-shot, Loss function=Modified Federated Loss2022.01 | 55.2 | 68.8 | — | |
| DeticBackbone=Swin-B, Supervision source=IN-L (ImageNet-L) image-level, Evaluation protocol=Zero-shot, Loss function=Modified Federated Loss2022.01 | 53 | 67.1 | — | |
| Box-SupervisedTraining dataset=LVIS-all, Supervision source=Box-supervised, Evaluation protocol=Zero-shot2022.01 | 46.2 | 61.7 | — | |
| Detictraining_dataset=LVIS, implementation=official2022.07 | — | — | 42.2 | |
| Detic-basetraining_dataset=LVIS, implementation=official2022.07 | — | — | 37.4 | |
| Ourstraining_dataset=LVIS, base_model=Detic-base2022.07 | — | — | 42.9 |