Is it possible to add negative samples(images without annotations) to mxnet version yolov3?

[shadyatscu]

In original yolov3 implementation, we could easily add negative samples(images without annotations) to training dataset. In mxnet version yolov3, i tried to do the same thing, e.g. for preset module 'yolo3_darknet53_coco' i have to create coco format data without annotation information. The coco data api of glouncvhttps://github.com/dmlc/gluon-cv/blob/bc77dcb42f0549d5389e6b7fece358e12c820a83/gluoncv/data/mscoco/detection.py#L33, there is an option: ' skip_empty : bool, default is True Whether skip images with no valid object. This should be True in training, otherwise it will cause undefined behavior.' what is the meaning of 'undefined behavior'? In cpp version yolov3 repo https://github.com/AlexeyAB/darknet it is said that: 'desirable that your training dataset include images with non-labeled objects that you do not want to detect - negative samples without bounded box (empty .txt files) - use as many images of negative samples as there are images with objects', has anybody try out this strategy? thanks a lot!

[ivanquirino]

+1

[chinakook]

This is important function to reduce false positive. mmdetection add this function to its base. So I suggest that gluon-cv can also add images without annotations to its detection base.

[github-actions]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.