facebookresearch / facebookresearch/co-tracker
Removing Queries as the tracking progresses
- Dominant language
- Jupyter Notebook
- Stars
- 5.1k
- Forks
- 389
- PR merge metrics
- No merged PRs in 30d
Description
Hey, I'm amazed at how well this works, however I have an issue with implementing this for my use case.
In my scenario, both the camera and the points to track are moving, this makes it fairly hard, I do realize that, but its still works well.
My issue is, that as the video progresses, points that have long been out of frame get "re-detected" in absurd or weird locations because they roughly match the color of the last tracked point.
In my opinion, it would be helpful to have a way to remove or mark queries as "do not track further", or any other way to stop tracking them. I tried adjusting the queries mid-processing, but this did not work, made obvious by this code:
[predictor.py](https://github.com/facebookresearch/co-tracker/blob/main/cotracker/predictor.py#L242)
showing it's only set on the first step.
Any ideas? Could I just modify the code to set queries even if it's not the first step, or will this break things?
Contributor guide
Research direction
Start in cotracker/predictor.py around line 242 and trace how queries are initialized and used after the first processing step. Reproduce the moving-camera case described in the issue, then define completion as allowing an individual query to stop being tracked without later re-detection at a new location.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python
- Domain
- computer-vision, machine-learning
- Issue type
- Feature
- Difficulty
- 4/5
- Estimated time
- 3-5 days
- Activity status
- Stale
- Clarity
- Mostly clear
- Newbie friendliness
- 28/100