Object Detection on MS-COCO 2017 (test-dev)
53.8mAP (50:95)Humble teacher
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| Humble teacherModel architecture=Cascade R-CNN, Backbone=ResNet-152, Training Data=MS-COCO train 2017 + MS-COCO unlabeled2021.06 | 53.8 | — | — | — | — | — | |
| R-101-P2HCT-SFLOPs (G)=138.5, Parameters (M)=49.9, Resolution=640x6402025.05 | 50.9 | 68.1 | 55.7 | — | — | — | |
| Base supervised modelModel architecture=Cascade R-CNN, Backbone=ResNet-152, Training Data=MS-COCO train 2017 + MS-COCO unlabeled2021.06 | 50.7 | — | — | — | — | — | |
| Adaptive Pairwise ErrorMethod=PAA*(800), Data Aug.=None, Epoch=36, Backbone=Res2Net-101-FPN-DCN, multi-scale training=true2022.07 | 50 | 67.7 | 53.9 | 31.8 | 55.3 | 66.7 | |
| Adaptive Pairwise ErrorMethod=PAA*(800), Data Aug.=None, Epoch=24, Backbone=ResNeXt-101-FPN-DCN2022.07 | 49 | 67.1 | 53 | 29.9 | 52.9 | 63.4 | |
| Generalize Focal LossMethod=ATSS(800), Data Aug.=None, Epoch=24, Backbone=ResNeXt-101-FPN-DCN, multi-scale training=true2022.07 | 48.2 | 67.4 | 52.6 | 29.2 | 51.7 | 60.2 | |
| R-50-P2HCT-NFLOPs (G)=71.9, Parameters (M)=26.0, Resolution=640x6402025.05 | 47.9 | 65.3 | 51.8 | — | — | — | |
| Focal LossMethod=ATSS(800), Data Aug.=None, Epoch=24, Backbone=ResNeXt-101-FPN-DCN, multi-scale training=true, Citation=[38]2022.07 | 47.7 | 66.5 | 51.9 | 29.7 | 50.8 | 59.4 | |
| G-FPN-D11FLOPs (G)=113.3, Parameters (M)=7.2*, Resolution=640x6402025.05 | 46.9 | — | — | — | — | — | |
| R-18-P2HCT-NFLOPs (G)=42.5, Parameters (M)=16.4, Resolution=640x6402025.05 | 46.3 | 63.7 | 49.6 | — | — | — | |
| G-FPN-D7FLOPs (G)=79.5, Parameters (M)=4.54*, Resolution=640x6402025.05 | 45.8 | — | — | — | — | — | |
| Distributional Ranking LossMethod=Retina(800), Data Aug.=None, Epoch=18, Backbone=ResNeXt-101-FPN, multi-scale training=true, multi-scale testing=true2022.07 | 44.7 | 63.8 | 48.7 | 28.2 | 47.4 | 56.2 | |
| R-101-A2-FPNFLOPs (G)=209.0*, Parameters (M)=62.5*, Resolution=640x6402025.05 | 42.8 | 65.2 | 47 | — | — | — | |
| R-101-AC-FPNFLOPs (G)=202.7*, Parameters (M)=62.5*, Resolution=640x6402025.05 | 42.4 | 65.1 | 46.2 | — | — | — | |
| Average Precision LossMethod=Retina(512), Data Aug.=SSD-style, Epoch=100, Backbone=ResNet-101-FPN, multi-scale training=true2022.07 | 42.1 | 63.5 | 46.4 | 25.6 | 45 | 53.9 | |
| R-50-AC-FPNFLOPs (G)=135.9*, Parameters (M)=44.7*, Resolution=640x6402025.05 | 40.4 | 63 | 44 | — | — | — | |
| R-101-AttnFPNFLOPs (G)=203.2*, Parameters (M)=64.7*, Resolution=640x6402025.05 | 40.2 | 62.5 | 43.6 | — | — | — | |
| R-101-Retina+SAFLOPs (G)=129.44, Parameters (M)=58.5, Resolution=640x6402025.05 | 40 | 60 | 42.9 | — | — | — | |
| R-50-A2-FPN-LiteFLOPs (G)=120.4*, Parameters (M)=44.5*, Resolution=640x6402025.05 | 39.8 | 62.3 | 43.4 | — | — | — | |
| R-50-AFPNFLOPs (G)=90.0, Parameters (M)=50.2, Resolution=640x6402025.05 | 39 | 57.6 | 42.1 | — | — | — | |
| R-50-AttnFPNFLOPs (G)=136.4*, Parameters (M)=46.9*, Resolution=640x6402025.05 | 38.4 | 61.1 | 41.9 | — | — | — | |
| Focal LossMethod=Retina(800), Data Aug.=None, Epoch=12, Backbone=ResNet-50-FPN, Citation=[19]2022.07 | 32.5 | 50.9 | 34.8 | 13.9 | 35.8 | 46.7 |