FSAF

This is an implementation of FSAF on keras and Tensorflow. The project is based on fizyr/keras-retinanet and fsaf branch of zccstig/mmdetection. Thanks for their hard work.

As the authors write, FASF module can be plugged into any single-shot detectors with FPN-like structure smoothly. I have also tried on yolo3. Anchor-free yolo3(with FSAF) gets a comparable performance with the anchor-based counterpart. But you don't need to pre-compute the anchor sizes any more. And it is much better and faster than the one based on retinanet.

Updates

• [03/05/2020] The author of the paper has released a new paper SAPD, which is based on FSAF. I have implemented it at xuannianz/SAPD.

Test

1. I trained on Pascal VOC2012 trainval.txt + Pascal VOC2007 train.txt, and validated on Pascal VOC2007 val.txt. There are 14041 images for training and 2510 images for validation.
2. The best evaluation results (score_threshold=0.05) on VOC2007 test are:

| backbone | mAP₅₀ | | ---- | ---- | | resnet50 | 0.7248 | | resnet101 | 0.7652 |

3. Pretrained models are here.
   baidu netdisk extract code: rbrr
   goole dirver

4. python3 inference.py to test your image by specifying image path and model path there.

Train

build dataset (Pascal VOC, other types please refer to fizyr/keras-retinanet)

• Download VOC2007 and VOC2012, copy all image files from VOC2007 to VOC2012.
• Append VOC2007 train.txt to VOC2012 trainval.txt.
• Overwrite VOC2012 val.txt by VOC2007 val.txt.

train

• python3 train.py --backbone resnet50 --gpu 0 --random-transform pascal datasets/VOC2012 to start training.

Evaluate

• python3 utils/eval.py to evaluate by specifying model path there.