search

facebookresearch / detr

End-to-End Object Detection with Transformers
Apache License 2.0
13.17k stars 2.39k forks source link

Question regarding self supervised object detection #185

Open thomasstats opened 3 years ago

thomasstats commented 3 years ago

I've been playing around with DETR for a little bit now and I've noticed through the attention visualizer in the notebook, that even objects that do not appear in coco or that the network were trained on, appear to have attention that is solely on that object. This means, from what I can determine, that DETR can actually detect objects even if it doesn't know what it is.

Have you guys been looking into exploring this space? I think using it as a more generic object detector, then combining that object detector with some kind of similarity, could split these objects found in attention into their own clusters, which could be very useful for things such as gathering annotations quickly.

Hatmm commented 3 years ago

Would you mind sharing the pics you tried on these experiences ?