Training Details

[lofrienger]

@Z-MU-Z
Hi, this is a very interesting and inspiring work!
However, can you please share more training details, such as dataset conversion, reward computation, and policy update?
Thank you for your time and for sharing this excellent work. Any insights you could provide would be greatly appreciated.

[Z-MU-Z]

Thanks for the question!

We’ve uploaded the trajectories and the step-wise masks of the dataset.

During training, the mask is applied online onto the image at each step (see the reference here: https://github.com/aim-uofa/SegAgent/blob/main/evaltools/model_loader.py#L361C12-L361C24
).

We only used SFT (no RL) for training.

For trajectory filtering, the reward is the IoU with the GT mask.