open-mmlab/mmdetection · info

gt_masks is already contained in results, so paste_by_box is

Error message

gt_masks is already contained in results, so paste_by_box is disabled.

What it means

Warning from CopyPaste transform (_get_gt_masks): results already contains 'gt_masks', so the paste_by_box mode (generating masks from boxes) is disabled and existing masks are returned unchanged.

Source

Thrown at mmdet/datasets/transforms/transforms.py:3081

    @cache_randomness
    def _get_selected_inds(self, num_bboxes: int) -> np.ndarray:
        max_num_pasted = min(num_bboxes + 1, self.max_num_pasted)
        num_pasted = np.random.randint(0, max_num_pasted)
        return np.random.choice(num_bboxes, size=num_pasted, replace=False)

    def get_gt_masks(self, results: dict) -> BitmapMasks:
        """Get gt_masks originally or generated based on bboxes.

        If gt_masks is not contained in results,
        it will be generated based on gt_bboxes.
        Args:
            results (dict): Result dict.
        Returns:
            BitmapMasks: gt_masks, originally or generated based on bboxes.
        """
        if results.get('gt_masks', None) is not None:
            if self.paste_by_box:
                warnings.warn('gt_masks is already contained in results, '
                              'so paste_by_box is disabled.')
            return results['gt_masks']
        else:
            if not self.paste_by_box:
                raise RuntimeError('results does not contain masks.')
            return results['gt_bboxes'].create_masks(results['img'].shape[:2])

    def _select_object(self, results: dict) -> dict:
        """Select some objects from the source results."""
        bboxes = results['gt_bboxes']
        labels = results['gt_bboxes_labels']
        masks = self.get_gt_masks(results)
        ignore_flags = results['gt_ignore_flags']

        selected_inds = self._get_selected_inds(bboxes.shape[0])

        selected_bboxes = bboxes[selected_inds]
        selected_labels = labels[selected_inds]

View on GitHub (pinned to cfd5d3a985)

Solutions

1. Set paste_by_box=False in the CopyPaste transform (masks are used anyway)
2. Remove the flag from config to silence the warning
3. Keep as-is if harmless; behavior is correct (uses real masks)

Example fix

# before
train_pipeline = [dict(type='CopyPaste', paste_by_box=True), ...]
# after
train_pipeline = [dict(type='CopyPaste', paste_by_box=False), ...]

if 'gt_masks' in results and copy_paste_cfg.get('paste_by_box', False):
    copy_paste_cfg['paste_by_box'] = False  # masks already exist

When it happens

Trigger: Configuring CopyPaste with paste_by_box=True on a pipeline where gt_masks already exist (segmentation annotations loaded).

Common situations: Instance segmentation configs reused with copy-paste augmentation; paste_by_box is a fallback for detection-only data, so it is redundant — and warned about — when masks are present.

Related errors

• dataset_meta or class names are not saved in the checkpoint'
• Checkpoint is not loaded, and the inference result is calcul
• weights is None, use COCO classes by default.
• palette does not exist, random is used by default. You can a
• Currently does not support saving datasample when return_dat

AI-assisted analysis of open-mmlab/mmdetection@cfd5d3a985 (2026-08-27). Data as JSON: /api/errors/a4509ea957d3264e. Report an issue: GitHub.