In your article in approach you said that common practice by the traditional methods is to post-process the optical flow using contextual information, such as median filtering and bilateral filtering.Are you use this filtering for you wight for pytorch or Caffe? I compare pytorch and tensorflow implantation on KITTI dataset and saw difference, that look like you used some filtering.
In your article in approach you said that common practice by the traditional methods is to post-process the optical flow using contextual information, such as median filtering and bilateral filtering.Are you use this filtering for you wight for pytorch or Caffe? I compare pytorch and tensorflow implantation on KITTI dataset and saw difference, that look like you used some filtering.