AruniRC / detectron-self-train

A PyTorch Detectron codebase for domain adaptation of object detectors.
MIT License
118 stars 21 forks source link

Will you open the source code for hard samples mining? #5

Open SpyderXu opened 5 years ago

SpyderXu commented 5 years ago

Will you open the source code for hard samples mining?

AruniRC commented 5 years ago

Our code was written for running on a cluster over multiple videos in parallel -- it is taking some time to clean this up for a public release. Apologies about this delay.

It is based off the MDNet tracker: https://github.com/HyeonseobNam/MDNet

Thank you for your patience.

Jason-cs18 commented 5 years ago

Thanks for your awesome works on self-learning. But I have questions about how to use MDNet to mine the hard positives because MDNet is a tracking algorithm for single object tracking and you need to provide bbs for multiple objects in this project. Besides this, I am confused about how to extract the hard bbs from tracking. From my perspective, MDNet is a detection-free tracking approach and would generate many bbs without detectors. Thus, did you use these new bbs by MDNet to filter all bbs (low confidence) by the detection?