Demo always predict 2 hands

[RoiPapo]

Hi, even on InterHand2.6M set that the model were trained on, when I try to visualize the prediction on single hand images using the demo.py ,
it shows both hands, I tried to debug and find out a number of hands evaluation but I couldn't it looks like the 2 hands case is hard coded on the demo code

(3d color was changed but nothing more)
thank you :)

[mks0601]

Hi,

Sorry about this. As you mentioned, the demo code is hard-coded for the two-hand cases.
One possible way to handle the single-hand case is checking confidence of hand boxes from the hand detection network.
Let me add that soon

[wefwwef4]

Hi,

Sorry about this. As you mentioned, the demo code is hard-coded for the two-hand cases. One possible way to handle the single-hand case is checking confidence of hand boxes from the hand detection network. Let me add that soon

hi, I still can't find where to cahnge confidence of hand boxes

[mks0601]

Hi, sorry for late.
You can use bbox confidence in here

InterWild/demo/demo.py

Line 112 in 0521504

hand_bbox_conf = float(out[h[0] + 'hand_bbox_conf'].cpu().numpy()[0]) # bbox confidence

[wefwwef4]

HI，Sometimes it is obvious that there is only the right hand, but the confidence of the left-handed bbx will suddenly increase. Is there a standard for judging?

[mks0601]

sorry that could be wrong predictions of the model.

[gmindflow]

Then would it be correct to understand that the DetectNet isn't clearly robust enough to distinguish "one hand" and "two hands extremely occluding another"?

Or is this only a problem with the demo code itself?

[mks0601]

I think your're right. There should be some room to improve DetectNet.