- Linux (tested on Ubuntu 16.04)
- mmdetection 2.0
- Python 3.7
- PyTorch 1.4
- Cython
- mmcv == 0.6.2
./tools/dist_train.sh <CONFIG_FILE> <GPU_NUM> --validate --work_dir <WORK_DIR>For example, our models use 3 RTX Titan GPUs for training
./tools/dist_train.sh configs/faster_rcnn_r50_drfpn_1x.py 3see more details at mmdetection.
| Backbone | detector | lr schedule | mAP(det) | mAP(mask) |
|---|---|---|---|---|
| ResNet-50 DRFPN | Faster R-CNN | 1x | 39.6 | - |
| ResNet-101 DRFPN | Faster R-CNN | 1x | 41.1 | - |
| ResNet-101 DRFPN | Faster R-CNN | 2x | 41.8 | - |
| ResNeXt-64x4d-101 DRFPN | Faster R-CNN | 1x | 43.4 | - |
| ResNet-50 DRFPN | Mask R-CNN | 1x | 40.0 | 36.3 |
| ResNet-101 DRFPN | Mask R-CNN | 1x | 41.9 | 37.8 |
| ResNet-101 DRFPN | Mask R-CNN | 2x | 42.7 | 38.4 |
| ResNet-50 DRFPN | Cascade Mask R-CNN | 1x | 43.0 | 37.7 |
| ResNet-101 DRFPN | Cascade Mask R-CNN | 1x | 44.5 | 38.8 |
| ResNet-50 DRFPN | RetinaNet | 1x | 38.0 | - |
| ResNet-50 DRFPN | FCOS | 1x | 37.9 | - |
| Backbone | detector | lr schedule | mAP(det) | mAP(mask) |
|---|---|---|---|---|
| ResNet-50 DRFPN | Faster R-CNN | 1x | 39.1 | - |
| ResNet-101 DRFPN | Faster R-CNN | 1x | 40.8 | - |
| ResNet-101 DRFPN | Faster R-CNN | 2x | 41.5 | - |
| ResNeXt-64x4d-101 DRFPN | Faster R-CNN | 1x | 43.0 | - |
| ResNet-50 DRFPN | Mask R-CNN | 1x | 39.7 | 35.9 |
| ResNet-101 DRFPN | Mask R-CNN | 1x | 41.7 | 37.4 |
| ResNet-101 DRFPN | Mask R-CNN | 2x | 42.4 | 38.5 |
| ResNet-50 DRFPN | Cascade Mask R-CNN | 1x | 42.8 | 37.4 |
| ResNet-101 DRFPN | Cascade Mask R-CNN | 1x | 44.1 | 38.3 |
| ResNet-50 DRFPN | RetinaNet | 1x | 37.6 | - |
| ResNet-50 DRFPN | FCOS | 1x | 37.6 | - |
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent