Closed sharifza closed 4 years ago
Hi, @sharifza . Yes, there is indeed a reason to take it out. In our original paper, as mentioned in README, we did not do the per-class NMS to filter the bounding boxes from object detection module, and thus the Co-NMS is essential to filter the highly-overlapped box-pairs from the candidates. However, in current implementation, we added the per-class NMS as Neural MotifNet since it improved the performance significantly, so there is no necessary to re-run the Co-NMS on top of the box-pairs any more. As a byproduct, this also saves time. Let me know you still have confuse.
Thanks for clarification (maybe you can explicitly mention this in the readme since its an important difference with the paper)
https://github.com/jwyang/graph-rcnn.pytorch/blob/be9d901a2cdecda3669eb780a3f437d6c3e061b9/lib/scene_parser/rcnn/modeling/relation_heads/relpn/relpn.py#L241
Hi! Based on the implementation in part of the repo, can I assume that your current implementation (and results) of RePN does not have the IoU-based NMS step described by Eq.3 of the paper? Was there a reason to take it out?