Method major

ext/spt/data/instance.py:159–222

Return the obj, count, and y of the majority instance in each cluster (i.e. the object with which it has the highest overlap). :param num_classes: int Number of classes in the dataset. Specifying `num_classes` allows identifying 'void' labels. By convention,

(
            self,
            num_classes: int = None
    )

        return self.obj.unique().numel()

    def major(
            self,
            num_classes: int = None
    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
        """Return the obj, count, and y of the majority instance in each
        cluster (i.e. the object with which it has the highest overlap).

        :param num_classes: int
            Number of classes in the dataset. Specifying `num_classes`
            allows identifying 'void' labels. By convention, we assume
            `y ∈ [0, self.num_classes-1]` ARE ALL VALID LABELS (i.e. not
            'ignored', 'void', 'unknown', etc), while `y < 0` AND
            `y >= self.num_classes` ARE VOID LABELS. Void data is dealt
            with following https://arxiv.org/abs/1801.00868 and
            https://arxiv.org/abs/1905.01220
        """

        # If `num_classes` was not passed, we set it to `y_max + 1`
        # (i.e. there are no 'void' objects)
        num_classes = num_classes if num_classes else self.y.max() + 1

        # Compute the cluster index for each overlap (i.e. each row in
        # self.values)
        cluster_idx = self.indices

        # Search the overlaps with void objects
        pair_is_void = (self.y < 0) | (self.y >= num_classes)

        # Search for the obj with the largest overlap, for each cluster
        x = torch.stack((self.count, self.count * ~pair_is_void)).T
        res = scatter_max(x, cluster_idx, dim=0)
        count = res[0][:, 0]
        argmax = res[1][:, 0]
        obj = self.obj[argmax]
        y = self.y[argmax]

        # If no cluster mainly overlaps with a void object, exit here
        is_major_void = (y < 0) | (y >= num_classes)
        if (~is_major_void).all():
            return obj, count, y

        # Otherwise, we need to find those clusters which overlap with
        # void, but with less than 50%. These clusters will not be
        # assigned to their main void cluster, but to their second-best
        # overlap. This way, only clusters with +50% void overlap will
        # be excluded from metrics computation, as defined in:
        # https://arxiv.org/abs/1801.00868

        # Search if any of the clusters assigned to a void object have
        # less than 50% void points
        total_count = scatter_sum(self.count, cluster_idx, dim=0)
        major_50_plus = (count / total_count) > 0.5
        if major_50_plus[is_major_void].all():
            return obj, count, y

        # Assign the clusters with less than 50% void overlap to their
        # second-best overlap
        count_no_void = res[0][:, 1]