Problems about computing the loss

[ChunmengLiu1]

Hi, thank you for your great effort! I feel confused about the loss computation.

1. For the Lcls loss, is the ground truth the class label of each image?
2. For the Lmask_dice and Lmask_bce, is the ground truth the binary mask without class information?
3. why did you generate the point labels? It shows as follows:

with torch.no_grad():
            # sample point_coords
            point_coords = get_uncertain_point_coords_with_randomness(
                src_masks,
                lambda logits: calculate_uncertainty(logits),
                self.num_points,
                self.oversample_ratio,
                self.importance_sample_ratio,
            )
            # get gt labels
            point_labels = point_sample(
                target_masks,
                point_coords,
                align_corners=False,
            ).squeeze(1)

4. You used prompt engineering. I read the code and found that you use the template to generate ov_classifier_weight in san.py lines 161-164 during training. Have you used prompt engineering anywhere else?
5. For inference, where did you use prompt engineering and class information? Is the class information used as known information for a single picture when inferring?
6. Can this work be considered as zero-shot semantic segmentation and Why?

These problems confused me a lot! I'm looking forward to your reply.