search

NVlabs / PoseCNN-PyTorch

PyTorch implementation of the PoseCNN framework
Other
195 stars 49 forks source link

Segmentation performance is too different compared with the paper. #1

Closed oorrppp2 closed 3 years ago

oorrppp2 commented 3 years ago

Firstly, I appreciate for your release of the pytorch code. I just succeeded running the code, and I get the clear predicted labels through feeding the example images in 'data/demo/*color.png' as input.

But when I run the script demo.sh with ycb video dataset image, I couldn't get the clear predicted labels. It was creating a label that looked like a superpixel for a background that is not a target object. It was different compared with the paper or YouTube (https://www.youtube.com/watch?v=B7I7R1GdzV8).

image

image

image

I run the demo.sh, and print out predicted labels. I wonder it is right.

How can I get the clear segmentation with YCB video dataset like yours? I used 'ycb_object', 'ycb_video' models, and all of them were bad.

Thank you! Regard,

yuxng commented 3 years ago

I am not sure what is wrong, but these results are not correct. Make sure you used the checkpoints in the folder 'ycb_video', and also check the script experiments/scripts/ycb_video_test.sh for testing on ycb video.