A question about double head implementation

[HB-X]

I noticed that the roi used in the classification branch and the regression branch are different. The roi used in the regression branch is larger than the one used in the classification branch because of a scaling factor added to the configuration file. What is the reason for this? Does this have a significant impact on the final detection accuracy? Thank you.

mmdetection/configs/double_heads/dh_faster_rcnn_r50_fpn_1x_coco.py

Line 5 in 9d3e162

reg_roi_scale_factor=1.3,

mmdetection/mmdet/models/roi_heads/double_roi_head.py

Line 21 in 9d3e162

bbox_reg_feats = self.bbox_roi_extractor(

[zachcoleman]

This idea appears to come from the official implementation here.

My best guess is: this gives the convolutional head more information rather than padding zeros or some other argument about additional spatial information.

[HB-X]

@zachcoleman Thank you for your help. In the officical implementation, I found the explanation given by the author.
wuyuebupt/doubleheadsrcnn#2 (comment)
Intuitively, I think this strategy has a significant impact on the detection accuracy. I will do a comparison experiment as soon as possible.

[HB-X]

I retrained double head by setting reg_roi_scale_factor=None, but the accuracy is much lower than the original one (39.3 AP vs 40.0 AP). The result demonstrates that the strategy of expanding roi does have a very significant impact on the detection accuracy. For fair comparison, I think it is better to specify this trick in configuration file.

[ZwwWayne]

This setting is part of the paper's novelty. But you are right, it would be better to add some comments to explain it.