主要提供了一种无监督的deep feature的提取方式
 
good point应该满足:
  1. they should be distributed more or less evenly throughout the image;
  2. have good repeatability between different view- points;
  3. be recognizable and distinguishable with descrip- tors;
  4. should not lie too densely.
可以认为是在superpoint上的改进
 
网络架构:
 GoodPoint: unsupervised learning of keypoint detection and description∗
 
train:
loss:
GoodPoint: unsupervised learning of keypoint detection and description∗GoodPoint: unsupervised learning of keypoint detection and description∗
 
首先构造gt(使用随机homograph+随机噪声派生出图像)
 
4.1 Keypoints loss
为32×32或16×16大小的每个区域选择一个关键点是基于这样的假设:关键点应该在整个图像中均匀分布,但不要太密集。
GoodPoint: unsupervised learning of keypoint detection and description∗
GoodPoint: unsupervised learning of keypoint detection and description∗
 
其中Lkeypoints 的loss 是这样计算:
GoodPoint: unsupervised learning of keypoint detection and description∗
如此自适应的解决detector问题(但是只是经过homograph没有办法解决金字塔呀?除非train数据中存在scale的大量的变化
 
4.2 Descriptor loss
GoodPoint: unsupervised learning of keypoint detection and description∗
 
所以Lgt表示的是detector层的heatmap的差异。
 
GoodPoint: unsupervised learning of keypoint detection and description∗idxgeom(j) ̸= idxdesc(j)∧distgeom(j) > 7. 不匹配的点的误差
 
GoodPoint: unsupervised learning of keypoint detection and description∗,类似Triples loss
结果:
GoodPoint: unsupervised learning of keypoint detection and description∗
比较superpoint提升不大 可能是比较新颖的不用标注数据吧
 
 
 
 
    

相关文章: