MIL陷入局部最优,检测到局部,无法完整的检测到物体。将instance划分为空间相关和类别相关的子集。在这些子集中定义一系列平滑的损失近似代替原损失函数,优化这些平滑损失。
C-MIL learns instance subsets, where the instances are spatially related, i.e., overlapping with each other, and class related, i.e., having similar object class scores.
C-MIL treats images as bags and image regions generated by an object proposal method [24,32] as instances
待解决的问题:
1) How to optimize the non-convex function
2) How to perform instance selection in the early training stages when the instance selector is not well trained.
to be continue ...
更完整的论文笔记[csdn]