search

zhaofang0627 / HPBTT

HPBTT-Pytorch
MIT License
36 stars 8 forks source link

What is the scale net? #2

Open xuxy09 opened 3 years ago

xuxy09 commented 3 years ago

I notice that there is a scale net in the code. However, I cannot find the corresponding descriptions in the paper. Could you please provide more explanations about this scale net?

Besides, I find that the scale net is not trained in the training code. I am also confused about this.

Thanks!

zhaofang0627 commented 3 years ago

The scale net is pretrained to scale the shape of SMPL body mesh for alignment with the input image because we do not use the camera parameters predicted by HMR. During training, the net is fixed like other SMPL parameters. You can also use the camera parameters predicted by HMR as we do on DeepFashion, it will produce similar results.

xuxy09 commented 3 years ago

Thanks for the reply!

In "train_market.py", the scale net seems only randomly initialized, but no pre-trained weights are loaded. Did I miss something here?

By the way, could you please let me know why you do not use the camera parameters of HMR for market dataset?

zhaofang0627 commented 3 years ago

The pretrained model of the scale net now can be loaded correctly by "--scale_model". You can find it in the models for Demo. Because the quality of images on Market-1501 is not good as training images used by HMR, we expect that the scale net trained dedicatedly for Market-1501 can work better than the HMR model. Actually, the result is slightly better (probably lifting 0.001 for SSIM).