search

JialeCao001 / D2Det

D2Det: Towards High Quality Object Detection and Instance Segmentation (CVPR2020)
https://openaccess.thecvf.com/content_CVPR_2020/papers/Cao_D2Det_Towards_High_Quality_Object_Detection_and_Instance_Segmentation_CVPR_2020_paper.pdf
MIT License
297 stars 86 forks source link

Some question about Discriminative RoI Pooling #12

Closed wangjue-wzq closed 3 years ago

wangjue-wzq commented 3 years ago

From your paper, I know that the Discriminative RoI pooling is specifically designed for classification. But after reading your code, I've seen the Discriminative RoI pooling used in SingleRoIExtractor. In other words, it operates on both classification and regression tasks. Am I right in understanding this? In addition, in the ablation experiment, DRP is significantly improved compared to AP of the baseline. Is the increase mainly derived from Defromable ROI pooling or Adaptive pooling? Look forward to your answers.

JialeCao001 commented 3 years ago

@wangjue-wzq Thanks for interest. Discriminative RoI pooling is specifically designed for classification, where Deformable ROI pooling has a good improvement. The regression uses ROI pooling instead of Discriminative RoI pooling. Please refer to the code.

https://github.com/JialeCao001/D2Det/blob/7bb9e5bc35fc6fad35d51910d012769a13f86de3/configs/D2Det/D2Det_detection_r50_fpn_2x.py#L53

wangjue-wzq commented 3 years ago

Thanks！