input image normalization

[khanhha]

Hello,

sorry for asking you too many questions.

I would like to know where the input images are actually normalized to ImageNet mean and standard deviation?

As far as I know, the code just subtracts the input image by the pixel_means = np.array([[[123.68, 116.78, 103.94]]])

The input heatmaps are in the range of [0, 255.0]

I can see two problems here

• the input value ranges are big >> 1.0, doe sit cause any gradient problem?
• the input value range of input image and input heatmaps are not on the same scale. Input images are subjected from the pixel_means, while the heatmap value ranges are in [0, 255].

Could you please describe a bit more about this kind of normalization? Or Did I miss something from your code?

Thanks!

[mks0601]

• the input value ranges are big >> 1.0, doe sit cause any gradient problem?
  -> No. Many works just use 0~255 pixel valued-images as the input of their networks.

• the input value range of input image and input heatmaps are not on the same scale. Input images are subjected from the pixel_means, while the heatmap value ranges are in [0, 255].
  -> The output scale does not have to be the same with that of the input.

[khanhha]

Thanks for your answer.

-> The output scale does not have to be the same as that of the input:
--> Sorry for the misunderstanding. I mean the scale value between two kinds of input: input image and input keypoint heatmaps; Image has a value range of [-126, 126], and input heatmap has a value of a range of [0, 255]. They are concatenated into a multi-channel input before passing into convolution operations, so I suppose that they should be normalized to the same scale.

Anyway, your normalization technique produces accurate results, so that means it works. Maybe it's just a matter of training convergence.

Best

[mks0601]

Although they are concatenated, their scales do not have to be in the same range. The learnable weights will consider the scale difference. I think you don't have to worry about that.

[khanhha]

Thank you very much