How is re_coco_gt calculated? Difference between re_coco_det and re_coco_gt

Hello,

I know that both of `re_coco_det` and `re_coco_gt` have the following two fields
- `img_feat`
- `img_bb`

And `img_bb` in `re_coco_gt` is ground-truth bounding boxes while in `re_coco_det`, it is predicted bounding boxes from Faster-RCNN. What about `img_feat`? How is `img_feat` calculated in `re_coco_gt` and `re_coco_det`, respectively?

Thanks for any input!

Start by locating the data-generation code for re_coco_gt and re_coco_det, then trace how Faster-RCNN produces the predicted boxes and image features. Compare both feature-generation paths and document what img_feat and img_bb contain in each dataset. Done means the calculation and the ground-truth-versus-detection difference are clearly explained.

Written by the indexing model from the issue text.