文章目录
- Bridging the Gap Between Anchor-based and Anchor-free Detection via Adaptive Training Sample Selection
- Probabilistic Anchor Assignment with IoU Prediction for Object Detection
- AutoAssign: Differentiable Label Assignment for Dense Object Detection
- OTA: Optimal Transport Assignment for Object Detection
Bridging the Gap Between Anchor-based and Anchor-free Detection via Adaptive Training Sample Selection
2020 CVPR
我在另外一篇博客中介绍过了,这里就简述一下。
每一个scale的特征图给一个固定尺寸的anchor。对一个gt:1)在每个特征图中找k个中心点距离最近的anchor,并计算IoU;2)对于所有scale的特征图拿到的anchor和它们的IoU,计算IoU的mean和std,IoU在mean+std以上的anchor设置为正。
ATSS起到了根据center来