2018ECCV文章,對於faster rcnn 的分類能力的思考:
most hard false positives result fromclassification instead of localization.
文中說,大部分的FP是分類錯誤導致的而不是迴歸(這個觀點我持保留意見)。
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.
總結爲三點思考:
1.分類和迴歸本身兩個任務的不同,然後目前的結構是兩者共享前面的特徵,這種方式不是最優的
2.多任務聯合訓練的loss可能導致單個任務陷入局部最優
3.不同尺度上的大感受野可能會導致在檢測小物體時有冗餘的特徵