search

facebookresearch / co-tracker

CoTracker is a model for tracking any point (pixel) on a video.
https://co-tracker.github.io/
Other
3.79k stars 249 forks source link

Evaluate on custom dataset #96

Open dat-nguyenvn opened 4 months ago

dat-nguyenvn commented 4 months ago

Now, my data has video frames + segmentation mask for each frame. My target: creating my dataset to evaluate Cotracker in my case study. Can you give me dataset structure example?

  1. On github you mentioned download dataset "TAP-Vid". Can you explain more?
  2. Or you can send me of "DAVIS First" test set that you used. My email: ah23975@bristol.ac.uk
  3. As I understand: inputs are Frames + segmentation mask + raft masks and the label (annotation) is list points track in every frame. Do I missing some thing? Please let me know your pre-processing.

Thanks Dat

nikitakaraevv commented 3 months ago

Hi @dat-nguyenvn,

  1. I think I meant that the python dataset class should be similar to TAP-Vid.
  2. You can download the TAP-Vid benchmark by following the instructions in this github repo: https://github.com/google-deepmind/tapnet
  3. I believe they also use an Optical Flow method to initialize trajectories for faster human annotation. They should have all the details in the paper.

Thanks, Nikita