search

yjh0410 / YOWOv2

The second generation of YOWO action detector.
MIT License
207 stars 32 forks source link

recognition problem #2

Closed snehashis1997 closed 1 year ago

snehashis1997 commented 1 year ago

is there any way so that I can integrate yowov2 into my current yolox detector module? take the box from yolox head, pass it to the yowov2 classification part, and classify that particular object

yjh0410 commented 1 year ago

@snehashis1997 YOWOv2, trained on AVA, can simultaneously locate the person in the current frame and identify the actions end-to-end. In my opinion, it seems that it may not be appropriate to assign the location person's bbox given by YOLOX head to YOWOv2 for classification. However, in my experience, YOWOv2 sometimes gives false positive predictions. For example, there are no people in some frames, but YOWOv2 still detects the target. Perhaps you can use YOLOX's head to suppress this false positive prediction. On the other hand, joint training of YOWOv2 and YOLOX to fully utilize YOLOX's powerful positioning capabilities may be a valuable way.

snehashis1997 commented 1 year ago

Thank you for your valuable response. How about using blank background images during training? as of now, I do not think it is possible. How to do that.

tiamo405 commented 3 weeks ago

Thank you for your valuable response. How about using blank background images during training? as of now, I do not think it is possible. How to do that.

det_conf = float(bbox[4]) is the probability of being human?