cvat
cvat copied to clipboard
TransT tracker integration
Hi, I've recently started to use CVAT to annotate videos of traffic scenes. Before coming across CVAT I briefly experimented with Supervisely and was very impressed by their semi-automatic video annotation with TransT . TransT is a single object tracking algorithm and was presented recently at CVPR21, a Pytorch implementation is available here. So I picked that up and integrated it into CVAT as a new tracking tool. If you find this useful I would like to open a pull request for this new feature.
a demo video showing the workflow and a comparison between TransT and SiamMask if you are interested: https://www.youtube.com/watch?v=_oT44BXMaJY
@dschoerk , could you please duplicate the feature request in the repo https://github.com/cvat-ai/cvat? The core team doesn't support the repo anymore.
Sure, I'll create an issue on the new repo.
Why was this one abandoned?
@dschoerk , thanks for the contribution. It looks like the tracker was integrated. I will close the issue: https://github.com/opencv/cvat/tree/develop/serverless/pytorch/dschoerk/transt/nuclio
Hi @dschoerk, thanks for the integration. However, the tracker seems quite slow and we have to keep clicking next frame in a video for it to work right? Seems counter-intuitive, and I wonder why not just use interpolation method and do it manually? Granted, I'm running on CPU, but will there be a difference even if I run on GPU?