evaluation/clips/README.md

Evaluation script for clips - Point-wise accuracy

clips_gt.json contains ground truth labels for all clips in the dataset.

Save your predicitons in a json file with this format:

{
  "<clip_name>": {
    "<frame number>": [<x-coordinate>, <y-coordinate>],
    "<frame number>": [<x-coordinate>, <y-coordinate>],
    ...
  },
  ...
}

Where clip_name should be same as the __key__ of the webDataset sample.

To evaluate use:

python eval.py --predictions PATH_TO_JSON