How does the leaderboard PSNR calculated?

[不爱吃猫的鱼]

I have calculated the PSNR at local machine, and below is my psnr calc code:

 mse = lambda x,y: (np.sum(np.square(y - x))) / float(x.size)

 psnr = lambda x,y: 10*math.log10(255. ** 2 / mse(x,y))

and I get about 31.9 PSNR, however on the leaderboard I only get 26.72.

What's the difference causing that?

[ltheis]

It looks like you are computing a PSNR value for each image, and then average those PSNR values.

However, we calculate a MSE across all pixels of all images, and the compute a PSNR from that.

[不爱吃猫的鱼]

I don't think it is because of it. because the avg mse of all images is 42.8738409156, then the corresponding PSNR is 10*math.log10(255. ** 2 / 42.87) = 31.81, there is still a gap between 26.72, could you public the evaluating calculation code of
PSNR and MS-SSIM? Thanks!

在 2018年4月11日星期三 UTC+8下午7:02:19，ltheis写道：

[ltheis]

I posted the PSNR code here: https://groups.google.com/forum/#!searchin/clic-2018/psnr$20code%7Csort:date/clic-2018/-DFK4w7U1wQ/vMLPMxoICwAJ

We use this MS-SSIM code (with default arguments): https://github.com/tensorflow/models/blob/master/research/compression/image_encoder/msssim.py#L125

[不爱吃猫的鱼]

thanks

在 2018年4月11日星期三 UTC+8下午8:59:20，ltheis写道：

[不爱吃猫的鱼]

I found the error , because I use uint8 dtype array, then the diff is always between 0~255, which reduced my mse. Hoping it is help for other ones!

在 2018年4月11日星期三 UTC+8下午6:58:46，不爱吃猫的鱼写道：