packnet-sfm
packnet-sfm copied to clipboard
BUG: SemiSupCompletionModel do not flip back the output['inv_depths_rgbd']
Bug appears when I train SemiSupCompletionModel. The DEPTH_PP_GT metric is much worse than DEPTH_GT of my re-trained model.
I check the traning code. In https://github.com/TRI-ML/packnet-sfm/blob/f59b1d615777a9987285a10e45b5d87b0369fa7d/packnet_sfm/models/SfmModel.py#L75 , I found if flip is True, you do flip back the output['inv_depths'] but not output['inv_depths_rgbd']. see: https://github.com/TRI-ML/packnet-sfm/blob/f59b1d615777a9987285a10e45b5d87b0369fa7d/packnet_sfm/models/model_utils.py#L127 , 'inv_depths_rgbd' key in not in the fliter_dict. As a result, it causes the model can not handle fliped input.
Thank you for pointing that out! And thank you for your PR, I will take a look and merge soon.