Been working on this tool for my PhD which involves training multi task vision models using various pre-trained models as inputs or pseudolabels in order to improve generalization. I work mostly on UAV datasets, but it should work okay on indoor scenes or self driving (at least Marigold and Mask2Former).

For example, this dataset was generated using this tool: https://huggingface.co/datasets/Meehai/dronescapes

I'm quite aggressively trying to "just get the nn.Module" from the public repos that other researchers put up in their overly convoluted frameworks. A simple `forward(rgb_input: torch.Tensor) -> torch.Tensor` is nice, having 100 imports from a generic framework that has versions incompatibilities with everything else is not.

PS: most mains are standalone runnable too, i.e. - https://gitlab.com/meehai/video-representations-extractor/-/... or - https://gitlab.com/meehai/video-representations-extractor/-/...