are the GECO2 training steps different from the GECO training process?

[Shwetha-Javagal]

Can I use the same GECO training script to train this model on my own object category to improve accuracy?

[jerpelhan]

Unlike GeCo, GeCo2 does not use pretraining. While the auxiliary loss is newly introduced in GeCo2, the overall training regime remains largely unchanged.

[Shwetha-Javagal]

Thanks for your response. I’m planning to train GECO2 on a custom dataset. Since training code isn’t released yet, could you please clarify:

1. Can GECO’s training pipeline (few-shot sampling + detection loss) be reused directly for GECO2?
2. For the auxiliary loss, is the “avg object size < 25 px” checked per image, and is it measured before or after resizing to 1024×1024?
3. Should the SAM2 Hiera backbone stay frozen for the full training?

[jerpelhan]

Sorry for the delay; I am currently occupied with other commitments. The training code will be released in early January, or at the latest by mid-January.

1. Yes.
2. After all processing is completed, as objects may appear small due to tiled augmentation. What matters is the effective object size as presented to the model.
3. Yes.