Codes for "Bridging the gap between one-to-many and one-to-one label assignment via NMS-aware alignment module". This project is based on mmdetection framework.
- MMDetection version 2.11.0.
- Please follow official installation guides.
# assume that you are under the root directory of this project,
# and you have activated your virtual environment if needed.
# and with COCO dataset in 'data/coco/'.
bash ./tools/dist_train.sh configs/NMS-aware/fcos_r50_fpn_o2m_o2o_2GCN_1x.py 2bash ./tools/dist_test.sh configs/NMS-aware/fcos_r50_fpn_o2m_o2o_2GCN_1x.py work_dirs/fcos_r50_fpn_o2m_o2o_2GCN_1x/epoch_12.pth 4 --eval bboxWe provide the following trained models. All models are trained with 16 images in a mini-batch. It's normal to observe ~0.2AP noise in all methods.
| Model | MS train | Lr schd | mAP | Config | Download |
|---|---|---|---|---|---|
| FCOS_R50_FPN_o2m_o2o_2GCN_1x | N | 1x | 39.7 | config | baidu |
| FCOS_R101_FPN_o2m_o2o_2GCN_1x | N | 1x | 42.0 | config | baidu |
| FCOS_R50_FPN_o2m_o2o_2GCN_3x | Y | 3x | 42.6 | config | baidu |
| FCOS_R101_FPN_o2m_o2o_2GCN_3x | Y | 3x | 44.6 | config | baidu |
| Faster_R50_FPN_o2m_o2o_2GCN_1x | N | 1x | 37.9 | config | baidu |
| Faster_R101_FPN_o2m_o2o_2GCN_1x | N | 1x | 39.8 | config | baidu |