I still have some questions about reproducing your result.
I use https://github.com/MILVLG/bottom-up-attention.pytorch to extract features, download r101'weight file(bua-caffe-frcn-r101_with_attributes.pth), and save the feature map from ’res4‘.（nms=0.3, score_thresh=0.1, min_bbox_num=10 and max_bbox_num=100）
All the feature maps are around 100G. However, you mentioned that your feature maps are around 300G in other issue.
I consider that i may make a mistake in features extraction so that i can't reproduce your result in the paper.
Thanks for the implementation of the paper.
I still have some questions about reproducing your result.
I use https://github.com/MILVLG/bottom-up-attention.pytorch to extract features, download r101'weight file(bua-caffe-frcn-r101_with_attributes.pth), and save the feature map from ’res4‘.（nms=0.3, score_thresh=0.1, min_bbox_num=10 and max_bbox_num=100）
All the feature maps are around 100G. However, you mentioned that your feature maps are around 300G in other issue.
I consider that i may make a mistake in features extraction so that i can't reproduce your result in the paper.
Can you check my method above?Thank you!