MIL陷入局部最優,檢測到局部,沒法完整的檢測到物體。將instance劃分爲空間相關和類別相關的子集。在這些子集中定義一系列平滑的損失近似代替原損失函數,優化這些平滑損失。app
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.函數
C-MIL treats images as bags and image regions generated by an object proposal method [24,32] as instances優化
待解決的問題:spa
1) How to optimize the non-convex function.net
2) How to perform instance selection in the early training stages when the instance selector is not well trained.3d
to be continue ...orm
更完整的論文筆記[csdn]blog