这是一篇CCF-C文章,简要描述一下思想:
目标检测任务的主动学习中的数据选择大多数都基于classification来,作者同时考虑了classification和localization。
主要提出了两个针对localization的方法:
1. Localization Tightness
如上图所示,计算了region proposal的输出框和最终预测框的IoU来表示Tightness:
T ( B 0 j ) = I o U ( B 0 j , R 0 j ) T(B_0^j)=IoU(B_0^j,R_0^j)T (B 0 j )=I o U (B 0 j ,R 0 j )
然后再结合classification,综合得出metric:
J ( B 0 j ) = ∣ T ( B 0 j ) + P m a x ( B 0 j ) − 1 ∣ J(B_0^j)=|T(B_0^j)+P_{max}(B_0^j)-1|J (B 0 j )=∣T (B 0 j )+P ma x (B 0 j )−1∣
这里我不敢苟同作者的metric,但是至少其思路挺好的
但是上述方法只能用于two-stage网络,region proposal也就是SS or RPN的output
; 2.Localization Stability
如图所示,通过增强图片的高斯噪声,评价噪声图片得到的框与原始图片得到的框的差距作为Stability:
S B ( B 0 j ) = ∑ n = 1 N IoU ( B 0 j , C n ( B 0 j ) ) N S_B\left(B_0^j\right)=\frac{\sum_{n=1}^N \operatorname{IoU}\left(B_0^j, C_n\left(B_0^j\right)\right)}{N}S B (B 0 j )=N ∑n =1 N IoU (B 0 j ,C n (B 0 j ))
然后再结合classification,综合得出metric:
S I ( I i ) = ∑ j = 1 M P max ( B 0 j ) S B ( B 0 j ) ∑ j = 1 M P max ( B 0 j ) S_I\left(I_i\right)=\frac{\sum_{j=1}^M P_{\max }\left(B_0^j\right) S_B\left(B_0^j\right)}{\sum_{j=1}^M P_{\max }\left(B_0^j\right)}S I (I i )=∑j =1 M P m a x (B 0 j )∑j =1 M P m a x (B 0 j )S B (B 0 j )
也就是根据预测的score加权评价。
这个metric同时适用于two-stage和one-stage网络
Experiment
T h e E n d . The\ End.T h e E n d .
Original: https://blog.csdn.net/qq_44537408/article/details/127162044
Author: 翁诗浩
Title: Localization-Aware Active Learning for Object Detection (ACCV)
原创文章受到原创版权保护。转载请注明出处:https://www.johngo689.com/681745/
转载文章受原作者版权保护。转载请注明原作者出处!