This repository reproduces "Zhu et al. Feature Selective Anchor-Free Module for Single-Shot Object Detection. CVPR, 2019." (FSAF) PDF in PyTorch. The implementation is based on MMDetection framework. All the codes for the FSAF model follow the original paper.

To use this repo, please follow README.md of MMDetection.

Train

• To train baseline (i.e., RetinaNet)

./tools/dist_train_retinanet_r50_400_050x.sh

• To train FSAF (w/o anchor-based (AB))

./tools/dist_train_fsaf_r50_400_050x.sh

Eval

For evaluation, pretrained model-weights should be located at "./models/here".

• To evaluate baseline (i.e., RetinaNet)

./tools/eval_retinanet_r50_400_050x.sh

• To evaluate FSAF (w/o anchor-based (AB))

./tools/eval_fsaf_r50_400_050x.sh

Below is benchmark results. All models are trained with an image-size of 400 and reduced LR-schedule for efficient experiments. Reproduced results show a similar aspect to the original paper (Table 1,2), demonstrating sanity of the implementation.

• Code reorganization is needed to be consistent with the style of MMDetection framework (current code is only written for fast prototyping).

• Ho-Deok Jang
• Email: hdjang@kaist.ac.kr
• Homepage: https://sites.google.com/view/hdjangcv