Method remove_void

ext/spt/data/instance.py:547–609

Return a new InstanceData with void clusters, objects and pairs removed. IMPORTANT: By convention, we assume `y ∈ [0, num_classes-1]` ARE ALL VALID LABELS (i.e. not 'void', 'ignored', 'unknown', etc), while `y < 0` AND `y >= num_classes` ARE VOID LABELS.

(
            self,
            num_classes: int
    )

        return is_a_void, is_pair_void, pair_cropped_count

    def remove_void(
            self,
            num_classes: int
    ) -> Tuple['InstanceData', torch.Tensor]:
        """Return a new InstanceData with void clusters, objects and
        pairs removed.

        IMPORTANT:
        By convention, we assume `y ∈ [0, num_classes-1]` ARE ALL
        VALID LABELS (i.e. not 'void', 'ignored', 'unknown', etc),
        while `y < 0` AND `y >= num_classes` ARE VOID LABELS.
        This applies to both `Data.y` and `Data.obj.y`.

        Points with 'void' labels are handled following the procedure
        proposed in:
          - https://arxiv.org/abs/1801.00868
          - https://arxiv.org/abs/1905.01220

        More precisely:
          - predictions (i.e. clusters here) containing more than 50% of
            'void' points are removed from the metrics computation
          - targets (i.e. objects here) containing more than 50% of
            'void' points are removed from the metrics computation
          - the remaining 'void' points are ignored when computing the
            prediction-target (i.e. cluster-object here) IoUs

        To this end, the present function returns:
          - `instance_data`: a new InstanceData object with all void
            clusters, objects, and pairs removed
          - `non_void_mask`: boolean mask spanning the clusters,
            indicating the clusters that were preserved in the
            `instance_data`. This mask can be used outside of this
            function to subsample cluster-wise information after
            void-removal

        NB: by construction, removing pairs in `pair_mask` from the
            InstanceData will also remove all target objects containing
            'void' points. Importantly, this assumes, however, that the
            raw instance annotations in the datasets are semantically
            pure: all annotated instances contain points of the same
            class. Said otherwise: IF AN INSTANCE CONTAINS A SINGLE
            'VOID' POINT, THEN ALL OF ITS POINTS ARE 'VOID'.
        """
        # Get the masks for indexing void clusters and pairs
        is_cluster_void, is_pair_void, pair_cropped_count = \
            self.search_void(num_classes)

        # Create a new InstanceData without void data
        idx = self.indices
        idx = idx[~is_pair_void]
        idx = consecutive_cluster(idx)[0]
        obj = self.obj[~is_pair_void]
        count = self.count[~is_pair_void]
        y = self.y[~is_pair_void]
        pair_cropped_count = pair_cropped_count[~is_pair_void]
        instance_data = self.__class__(idx, obj, count, y, dense=True)

        # Save the pair_cropped_count in the new InstanceData. This will