Regarding the issue of EfficientSAM of checkpoints about efficientsam

[yformer]

@feivellau, thanks for your interest! We finetuned the entire model including encoder and decoder. I guess it may not work with efficientsam encoder and original sam decoder. Maybe you can give it a try?

[feivellau]

@feivellau, thanks for your interest! We finetuned the entire model including encoder and decoder. I guess it may not work with efficientsam encoder and original sam decoder. Maybe you can give it a try?

@yformer, I have tried it and it doesn't work properly. Furthermore, I have found that the results are very poor when setting multimask_output=False to get a single mask.

[yformer]

@feivellau, the entire model is finetuned. It may not work properly with efficientsam encoder and original sam decoder. For a single mask, you can set multimask_output=True and then obtain the single mask from output with single_mask = predicted_logits[0, 0, 0, :, :] >= 0.

[feivellau]

@yformer.

if multimask_output and self.num_multimask_outputs > 1:
    return masks[:, 1:, :], iou_pred[:, 1:]
else:
    return masks[:, :1, :], iou_pred[:, :1]

I'm just trying to follow what the original SAM paper said about a couple different mask outputs. But the results masks[:, :1, :], iou_pred[:, :1] in EfficientSAM are not good. And masks[:, 1:, :], iou_pred[:, 1:] needs to be sorted to determine the best output

[feivellau]

The same issue as Expose multimask_output in EfficientSam.forward has some examples.