About KPD about udp-pose HOT 5 CLOSED

@MaxChu719 Please refer to the latest verson of UDP in arxiv for the explaination of this setting. https://arxiv.org/pdf/1911.07524.pdf

from udp-pose.

@HuangJunJie2017 I have read through the paper and the only place i found about KPD is in section 3.2.2 (Combined classification and regression format), where it sais "...r is a hyperparameter referring the radius of the area classified as positive...". But the KPD in the implementation, as i understand, not only define the radius of interest of the offset but also it scale down the magnitude of the offset. So, KPD also served as the inverse weight of the offset loss. Am i misundersand something here?

from udp-pose.

@MaxChu719

but also it scale down the magnitude of the offset.

Well, this is another question
The initial purpose of scaling down the magnitude of the offset is just empirically making the offset to distribute within interval [0,1). It indeed scales down the offset loss in practice and, as i guess, should affect the final performance.
Anyway, the implement of this paradigm is not the main concern in this paper, as far as it is unbiased. Some details are not considered thoroughly.

from udp-pose.

@HuangJunJie2017 Thanks for the quick reply. I see your point that it is to normalize the offset interval to [0, 1). But you can also interpret it as the weight of the (unnormalized) offset loss. The reason why KDP has a pretty significant result is quite surprising. Maybe it is just a hyperparameter that happens to be quite good for COCO dataset. Anyway, thanks for the explanation and your good work.

from udp-pose.

@HuangJunJie2017 Btw, in my original, question i said "But in the implementation, it also controls the slope (and intercept) of the two offset heatmaps (seel below code)", i admit that i am not saying in very clear. It actually just mean that you also scale the offset loss...

from udp-pose.