YOLOF
Tensorflow implement YOLOF(You Only Look One Level Feaature)
Supports and Results
- GIOU.
- Smooth L1 loss.
- Tensorflow pretrained resnetmodel zoo
- Tensorboard visualize detection results When training.
- Best Result in Pascal VOC(0.66, train and test in VOC07).I guess in COCO will get a better performance?
- I have to say this result is not so good as RetinaNet and FCOS, but the author did not provide their result in VOC, So I have no reference to make this code perform better. If anyone has a better result,I'll appreciate it if you share your experience.
Main difficulty
- GIOU did not converge! So I chose Smooth L1 loss and this code support this two methods to regress boxes. If you find some Error about GIOU in my code or you can find a way make it converge, Welcome to contact me!
- Larger resolution may reach a better results, but it decrese the inference speed largely.
Train/eval/demo
- If you want to train your own dataset, You can refer FCOS because they have similar code structure or you can contact me.
