search

xingyizhou / CenterNet

Object detection, 3D detection, and pose estimation using center point detection:
MIT License
7.2k stars 1.92k forks source link

diameter of gaussian generated heatmap #842

Open kakusikun opened 3 years ago

kakusikun commented 3 years ago

https://github.com/xingyizhou/CenterNet/blob/2b7692c377c6686fb35e473dac2de6105eed62c6/src/lib/utils/image.py#L127

the diameter is always an odd number which makes the maximum of heatmap generated from gaussian is ~0.7

but this generated heatmap is used to train the objectness in the heatmap which is probability

is the trick used to regularization or something?

xingyizhou commented 3 years ago

Thank you for your interest in our work. The gaussian heatmap code is simply copied from CornerNet. I think the diameter just make the rendering faster. The peak of the heatmap should always be 1.

ztb520 commented 1 year ago

Thank you for your interest in our work. The gaussian heatmap code is simply copied from CornerNet. I think the diameter just make the rendering faster. The peak of the heatmap should always be 1.

How does a detection task generate a heatmap?

JasOleander commented 2 months ago

https://github.com/xingyizhou/CenterNet/blob/2b7692c377c6686fb35e473dac2de6105eed62c6/src/lib/utils/image.py#L127

the diameter is always an odd number which makes the maximum of heatmap generated from gaussian is ~0.7

but this generated heatmap is used to train the objectness in the heatmap which is probability

is the trick used to regularization or something?

when you generate heatmap, how much time taken for 128 batch size?