1.将本项目中mmdet/model/neck中的nfpn.py以及__init__.py文件复制到mmdetection/mmdet/neck下,替换__init__.py文件
python tools/train.py config/faster_rcnn/nfpn.py
2.将本项目中mmdet/model/detectors中的twostage.py文件替换到mmdetection/mmdet/detectors下的twostage.py
3.将本项目中mmdet/model/roi_heads中的standard_roi_head.py文件复替换mmdetection/mmdet/roi_heads中的standard_roi_head
4.将本项目中mmdet/model/roi_heads/roi_extractors文件夹替换复替换mmdetection/mmdet/roi_heads中的/roi_extractors文件夹
表格 1 在Lisa数据集上的实验
Method | mAP | AP50 | AP75 | APS | APM | APL | backbone |
---|---|---|---|---|---|---|---|
Faster RCNN | 65.2 | 76.1 | 74.8 | 66.8 | 71.0 | 70.5 | Res50-FPN |
BFP | 63.9 | 75.9 | 74.0 | 63.4 | 71.2 | 77.9 | ResN50-FPN |
BiFPN | 58.4 | 70.9 | 69.8 | 55.4 | 65.1 | 83.5 | Res50-FPN |
BiFPN*2 | 49.2 | 59.9 | 58.4 | 49.1 | 56.6 | 78.5 | Res50-FPN |
NFPN(ours) | 66.1 | 77.5 | 75.8 | 65.9 | 72.1 | 75.7 | Res50-FPN |
MRoIE + NFPN(ours) | 67.4 | 78.8 | 77.5 | 68.0 | 74.3 | 76.0 | Res50-FPN |
NFPN*2+ Double ROI(ours) | 66.0 | 78.1 | 76.4 | 66.2 | 72.5 | 80.7 | Res50-FPN |
NFPN +RFP + Double ROI(ours) | 67.8 | 79.0 | 77.7 | 68.4 | 74.0 | 81.1 | Res50-FPN |
表格 2 在 VOC07+12上的实验
Method | mAP | backbone |
---|---|---|
Faster RCNN | 73.2 | VGG-16 |
Faster RCNN | 79.5 | Res-50 |
NFPN(ours) | 81.3 | Res-50 |
Faster RCNN | 76.4 | Res-101 |
NFPN(ours) | 82.4 | Res-101 |