Keyword [YOLOv3]
Redmon J, Farhadi A. Yolov3: An incremental improvement[J]. arXiv preprint arXiv:1804.02767, 2018.
1. Overview
In this paper, it proposes YOLOv3 for detection.
- Chnage hierarchical classification into multi-label classification.
- Predict boxes at 3 different scales. (3 boxes each scale).
- Based on DarkNet-53.
- Still multi-scale training, data augmentation, BN.
2. BBox Prediction
1) Predict 3 boxes at each scale $N \times N x [3 * (4 + 1+ 80)]$.
2) Scale means feature comes from different layers. (Shallow layer is good for small obj)
3) Use k-means find 9 clusters, then divide up the clusters evenly across scales. The 9 clusters On COCO are:
$(10 \times 13),(16 \times 30),(33 \times 23),(30 \times 61),(62 \times 45),$
$(59 \times 119),(116 \times 90),(156 \times 198),(373 \times 326)$