December 2020
tl;dr: A bag of tricks to train YOLOv3.
Overall impression
This paper and YOLOv4 both starts from YOLOv3 but adopts different methods. YOLOv4 explores extensively recent advances in backbones and data augmentation, while PP-YOLO adopts more training tricks. Their improvements are orthogonal.
The paper is more like a cookbook/recipe, and the focus is how to stack effective tricks that hardly affect efficiency to get better performance.
Key ideas
- Bag of training tricks
- Larger batch
- EMA of weight
- Dropblock (structured dropout) @ FPN
- IoU Loss in separate branch
- IoU Aware: IoU guided NMS
- Grid sensitive: introduced by YOLOv4. This helps the prediction after sigmoid to get to 0 or 1 position exactly, at grid boundary.
- CoordConv
- Matrix-NMS proposed by SOLOv2
- SPP: efficiently boosts receptive field.
Technical details
- Summary of technical details
Notes