- All models were trained on
coco_2017_train, and tested on thecoco_2017_val, for your conveinience of evaluation and comparison. In our paper, the numbers are obtained from test-dev. - To balance accuracy and training time when using InstaBoost, models released in this page are all trained for 48 Epochs. Other training and testing configs are strictly following the original framework.
| Framework | Network | Backbone | Lr schd | box AP | mask AP | Download |
|---|---|---|---|---|---|---|
| Detectron | Mask R-CNN | R-50-FPN | 4x | 40.3(orig:37.7) | 35.7(orig:33.7) | Baidu / Google |
| Detectron | Mask R-CNN | R-101-FPN | 4x | 42.8(orig:40.0) | 37.5(orig:35.9) | Baidu / Google |
| MMdetection | Mask R-CNN | R-50-FPN | 4x | 40.0(orig:37.3) | 36.2(orig:34.2) | Baidu / Google |
| MMdetection | Mask R-CNN | R-101-FPN | 4x | 42.1(orig:39.4) | 37.8(orig:35.9) | Baidu / Google |
| MMdetection | Mask R-CNN | X-101-64x4d-FPN | 4x | 44.5(orig:42.1) | 39.5(orig:38.0) | Baidu / Google |
| MMdetection | Cascade R-CNN | R-101-FPN | 4x | 45.4(orig:42.6) | 39.2(orig:37.0) | Baidu / Google |
| MMdetection | Cascade R-CNN | X-101-64x4d-FPN | 4x | 47.2(orig:45.4) | 40.4(orig:39.1) | Baidu / Google |
| MMdetection | SSD | VGG16-512 | 360e | 30.3(orig:29.3) | - | Baidu / Google |