2018ECCV文章,对于faster rcnn 的分类能力的思考:
most hard false positives result fromclassification instead of localization.
文中说,大部分的FP是分类错误导致的而不是回归(这个观点我持保留意见)。
We conjecture that:
(1) Shared feature representation is not optimal due to the mismatched goals of fea-
ture learning for classification and localization;
(2) multi-task learning helps, yet optimization of the multi-task loss may result in sub-optimal for individual tasks;
(3) large receptive field for different scales leads to redundant context information for small objects.
总结为三点思考:
1.分类和回归本身两个任务的不同,然后目前的结构是两者共享前面的特征,这种方式不是最优的
2.多任务联合训练的loss可能导致单个任务陷入局部最优
3.不同尺度上的大感受野可能会导致在检测小物体时有冗余的特征