Difference between YOLO and JDE

[wongkinyiu]

Difference between YOLO and JDE

[WongKinYiu]

yolo enable one box belongs to multiple classes, so it use sigmoid for decoding.
jde extract reid feature, an unique object must have only one id, so it use softmax for decoding.
it is better to take a look compute_loss function to understand the reason.

[WongKinYiu]

Sorry about that I can not understand your question.

[IgnacioTsai]

Sorry about that I can not understand your question.

No, I am the one should say sorry duo to weak expression.

My main question is "Does JDELayer is another prediction method of YOLO"?

The question what I means is the code from yolor/models/models.py , I saw that there is an hidden layer "JDELayer".
The keyword of "JDE" in YOLOR paper means "joint detection and embedding". I do not know what it means, but it looks like the structure of YOLOLayer, so that I compare with both code function in model.py.

I saw the only difference at the bottom of forward function.

YOLOLayer
io = p.sigmoid() io[..., :2] = (io[..., :2] * 2. - 0.5 + self.grid) io[..., 2:4] = (io[..., 2:4] * 2) ** 2 * self.anchor_wh io[..., :4] *= self.stride

JDELayer
io = p.clone() # inference output io[..., :2] = torch.sigmoid(io[..., :2]) * 2. - 0.5 + self.grid # xy io[..., 2:4] = (torch.sigmoid(io[..., 2:4]) * 2) ** 2 * self.anchor_wh # wh yolo method io[..., :4] *= self.stride io[..., 4:] = F.softmax(io[..., 4:])

It looks like different anchor predictions method, but I did not see any "jde" section in yolor relatived cfg.
That why I left this message.

By the way, thanks for reply.

[IgnacioTsai]

I realize how it works due to your explanation. Thanks.