For ImageNet localization, the authors observed that images usually contain one dominant object and RPN proposals overlap heavily, producing very similar RoI-pooled features. They argue that image-centric Fast R-CNN therefore yields low-variation stochastic samples, and use the original RoI-centric R-CNN instead. · CiteArk
For ImageNet localization, the authors observed that images usually contain one dominant object and RPN proposals overlap heavily, producing very similar RoI-pooled features. They argue that image-centric Fast R-CNN therefore yields low-variation stochastic samples, and use the original RoI-centric R-CNN instead.
来源：paper:PDF page 12, ImageNet Localization paragraph following Table 13
