about the size of tgt_pose when inference

[gouchaonijiao]

Hello! Your work is very impressive,but I wonder why the image and tgt_tensor are scaled to 256*256? And how I get other size of tgt_tensor like (256,512). Wish your reply!

[ankanbhunia]

You can get the other sizes using this code: #27 (comment).

You may also try resizing the existing pose files to your desired size (using Image.NEAREST in PIL library). Not sure though if it would work.

[gouchaonijiao]

Thanks for your reply！ But the problem is when change the shape of tgt_tensor(the parameter of "get_tgt_tensor.py"), it seems weird.
I've also tried to extract openpose coordinate in size of (512,256),but was still not work.
Or is there any method to transform openpose(image,not coordinate) into tgt_tensor since the image is visualized by the first 3 channels of tgt_tensor?

[ankanbhunia]

fixed one bug in the code. You can try it now. #27 (comment).

[gouchaonijiao]

Thanks for your answer,I'll adjust the code into my desired size then. Thanks again!

[gouchaonijiao]

Hello! Sorry to bother you again. I wonder know why you resize the image into (256,256) first when inference.? The generated image will undergo deformation or can you tell me which image size produces the best results?

[ankanbhunia]

Because the model was trained with (256, 256)-resized images, the resizing does not have any effect on the final output. Also, at the end, you resize the output back to its original dimension.

[gouchaonijiao]

I sincerely appreciate your patient answer and wish you complete more outstanding work in the future!