Face categorisation by entity

[mujavidb]

I can see in your evaluation code that for each detected entity in a video, you capture the face data for each frame, and then you pass these into the ASD detector on an entity by entity level.

How are you categorising faces by entity? I can see face detection is done with SF3D. But when load_visual is called the faces are already segmented by entity. So, at inference, each iteration of the data loader val_loader is done on a face-by-face level. It is unclear how categorisation is done by entity in this example code.

[Junhua-Liao]

Thank you for your interest in our work. The AVA dataset has provided entity information for each face. As for the Columbia dataset, you can refer to the 'track_shot' function in line 125 of the Columbia_test.py file.

[mujavidb]

Very helpful thanks.