Imbalanced examples in object detection

这周看了一篇《Learning a Unified Sample Weighting Network for Object Detection》.分析的是Imbalance between positive and negative examples.由此,找了其余的一些cvpr2020paper以及这里的经典methods.

Introduction:

Focal Loss in RetinaNet

FAIR Kaiming he组,ICCV 2017 best student paper

Other Paper in CVPR2020

key: 利用定义的值cleanliness来描述candidate proposal,通过localization accuracy和classification confidence的加权平均来计算;
通过这个cleanliness来动态调整不同candidate proposal在训练中的权重

key: not only ‘hard’ example matters, but ‘easy’ one. jointly re-weights classification and regression loss.

image-20201122145912027
$m_{i}^{reg}$和$m_{i}^{cls}$利用高斯分布推出,并进行正则化

key: ‘hard’ examples are not as import as ‘IOU-highest’ positive samples or ‘score-highest’ negative samples. Sort IOU and Scores in hierarchy and adjust weights dynamicly.

\[L_{cls} = \sum_{i=1}^{n}{w_{i}^{'}CE1} + \sum_{j=1}^{m}{w_{j}^{'}CE1} \\ w_{i}^{'}=w_{i} \frac{ \sum_{i=1}^{n}CE1}{ \sum_{i=1}^{n}{w_{i}}CE1} \\ w_{j}^{'}=w_{j} \frac{ \sum_{j=1}^{m}CE1}{ \sum_{j=1}^{m}{w_{j}}CE1} \\ tips:改变了权重,没有改变loss的总和\]

key: reducing rare category’s negative samples from easy category(positive samples in easy category will be regarded as negative in rare)

The same position with $(1-p_{t})^{\gamma}$ in focal loss, $w_{j}=1-E(r)T_{\lambda}(f_{j})(1-y_{i})$,;对$E(r)=1$,表示r是foreground类别,0则表示background;$T_{\lambda}(f_{j})=1$表示类别为rare,0表示为easy;$1-y_{j}=1$表示对当前类别为负样本negative,0表示positive; 因此,$W_{J}=0$表示,这个proposal被分为rare类别的负样本,也就是之前提到的easy类别的正样本,需要将这部分negative训练时产生的影响去除。