Use multiple (negative) points for training/inference

[ben-jy]

Hi,
Thanks for sharing your work.
I'm not very experienced with MMDetection or the OpenMMLab framework, therefore I’ve spent some time trying to understand how the project is structured.
My ultimate goal is to use the model with multiple point inputs as prompts instead of a single centroid point. I’d also like to evaluate the impact of adding negative points during inference (using point_labels attribute of SAM predict method).
At the moment, I’m not sure where this could be implemented or whether the current code already supports it. If not, do you think it would be possible to run the model outside of MMDetection, directly in PyTorch ?
I would greatly appreciated any guidance or pointers :-).

[ghafoorharis]

same painpoint ~

[piringhologram]

I was trying to run inference with the model path, without any groundtruth unlike the minimal working example. Could anyone guide me on how to do this? I have very minimal experience and is having some difficulty. would appreciate it :)