search

wwdkl / AABO

Implementation for ECCV 2020 paper: AABO: Adaptive Anchor Box Optimization for Object Detection via Bayesian Sub-sampling.
30 stars 5 forks source link

FPN anchor 分布 #1

Open risemeup opened 3 years ago

risemeup commented 3 years ago

你好， 我阅读了您的论文，关于论文fig2，coco数据集anchor分布我画出的图跟您的有一些区别。 分层log10轴 分层log10轴 请问您是画的coco的哪部分数据？还是抽样了一部分图片？ 希望大佬指点一下，万分感谢。

wwdkl commented 3 years ago

您好～ 关于您提到的Fig2 bbox在FPN不同层的分布问题，我们认为，受FPN不同层的不同感受野影响，较浅的那些层应当对scale更小的bbox负责，较深的层应当对scale更大的bbox负责，因此我们按照FPN不同层的感受野划分了每一层主要注意的bbox范围，之后又统计了这些bbox在ratio维度上的分布情况，是一种相对简单粗略的划分。希望能解决您的疑问。 感谢您对我们的工作感兴趣！祝好～

risemeup commented 3 years ago

thank you for your reply！ I have anther question.how to get the search space for different feature map? Your paper said

The regions within 5 red rectangles and the anchor distribution bounds represent the search space for each feature map in FPN QQ截图20201109200755

but how were the 5 rectangles designed?I did't find detailed explanation in the paper. By the way,will you public you code in the future?Waiting for your reply!