HAL-42/FMA-WSSS

upload the files to OneDrive, Google Drive, Dropbox?

ProjectDisR opened this issue · 7 comments

ProjectDisR commented
upload the files to OneDrive, Google Drive, Dropbox?
shakeebmurtaza commented

These links are expired. Please provide updated links.

Devin-Pi commented

Thanks for your work! But these links are expired. Could you Please provide updated links? Thank you!

J-Lam0 commented

hi, may I ask the graphics memory for generate SAM-based quasi-superpixel? I use 3090 24GB, need to decrease '-b' '-w ' for some complex images, but i think decrease these two parameters will reduce quasi-superpixel's quality.
could you please provied the quasi-superpixel?

HAL-42 commented

These two parameters do not affect performance.

Also try modify src/libs/sam/custom_sam/sam_auto.py with following _mask_intersection_areas implementation. This version significantly reduces the memory usage:

    @staticmethod
    def _mask_intersection_areas(mask1: torch.Tensor, mask2: torch.Tensor, einsum: bool=True) -> torch.Tensor:
        """Computes the intersection area between two masks.

        Args:
            mask1: Binary masks of shape (N1, H, W).
            mask2: Binary masks of shape (N2, H, W).
            einsum: 是否使用einsum计算。

        Returns:
            A tensor of shape (N1, N2) containing the intersection area between the two masks.
        """
        if not einsum:
            # * 计算两个mask的交集。
            intersection = mask1[:, None, :, :] & mask2[None, :, :, :]  # (N1, N2, H, W)

            # * 计算交集的面积。
            intersection_areas = intersection.sum(dim=(2, 3), dtype=torch.long)  # (N1, N2)
        else:
            intersection_areas = torch.einsum('nij,mij->nm',
                                              mask1.to(torch.float32), mask2.to(torch.float32)).to(torch.long)

        return intersection_areas

    @staticmethod
    def _chunk_mask_intersection_areas(mask1: torch.Tensor, mask2: torch.Tensor, chunksize: int=0) -> torch.Tensor:
        """Computes the intersection area between two masks.

        Args:
            mask1: Binary masks of shape (N1, H, W).
            mask2: Binary masks of shape (N2, H, W).
            chunksize: Chunk size to use for the computation.

        Returns:
            A tensor of shape (N1, N2) containing the intersection area between the two masks.
        """
        if chunksize == 0:
            return SamAuto._mask_intersection_areas(mask1, mask2)

        ret = []

        for i in range(0, mask1.shape[0], chunksize):
            chunk = mask1[i:i + chunksize, :, :]  # (chunksize, H, W)
            ret.append(SamAuto._mask_intersection_areas(chunk, mask2))  # (chunksize, N2)

        ret = torch.cat(ret, dim=0)  # (N1, N2)

        return ret
kuongan commented

the link is expired. Could you provide updated links? Thank you