would the focus mask branch downgrade performance of original detector in single-scale training and inference

[yaoqi-zd]

thanks for your impressive work, I would like to ask whether the focus mask branch will downgrade the performance of the detector in single scale train/inference setting ?

since the RCNN and focus branch are all built upon conv5 feature, the gradient from RCNN branch tends to highlight region of all objects, while the gradient of focus branch tends to hightlight only small object region, which seems a conflict and may bring negative impact on RCNN prediction.