What causes "palette does not match classes as metainfo is {self._metainfo}." in open-mmlab/mmdetection?

Set palette='random' (or omit it) to let the dataset generate colors automatically

open-mmlab/mmdetection · error · ValueError

palette does not match classes as metainfo is {self._metainf

Error message

palette does not match classes as metainfo is {self._metainfo}.

What it means

BaseSemanticSegmentationDataset._update_palette validates the palette against metainfo classes after label mapping. A palette longer than classes is allowed (extra entries ignored after mapping), but a palette shorter than the (mapped) number of classes raises this ValueError showing the full metainfo. Palettes can be a list of RGB tuples or an mmcv ColorStatus-like string ('random'/'none'), and list palettes must cover the classes.

Source

Thrown at mmdet/datasets/base_semseg_dataset.py:224

            np.random.seed(42)
            # random palette
            new_palette = np.random.randint(
                0, 255, size=(len(classes), 3)).tolist()
            np.random.set_state(state)
        elif len(palette) >= len(classes) and self.label_map is not None:
            new_palette = []
            # return subset of palette
            for old_id, new_id in sorted(
                    self.label_map.items(), key=lambda x: x[1]):
                # 0 is background
                if new_id != 0:
                    new_palette.append(palette[old_id])
            new_palette = type(palette)(new_palette)
        elif len(palette) >= len(classes):
            # Allow palette length is greater than classes.
            return palette
        else:
            raise ValueError('palette does not match classes '
                             f'as metainfo is {self._metainfo}.')
        return new_palette

    def load_data_list(self) -> List[dict]:
        """Load annotation from directory or annotation file.

        Returns:
            list[dict]: All data info of dataset.
        """
        data_list = []
        img_dir = self.data_prefix.get('img_path', None)
        ann_dir = self.data_prefix.get('seg_map_path', None)
        if not osp.isdir(self.ann_file) and self.ann_file:
            assert osp.isfile(self.ann_file), \
                f'Failed to load `ann_file` {self.ann_file}'
            lines = mmengine.list_from_file(
                self.ann_file, backend_args=self.backend_args)
            for line in lines:

View on GitHub (pinned to cfd5d3a985)

Solutions

Set palette='random' (or omit it) to let the dataset generate colors automatically
Provide one RGB tuple per class: len(palette) >= len(classes) after label mapping — e.g., copy the 3-length list to the needed size or take a slice of mmcv's palette constants
If subsetting classes, also shrink/reorder the palette to match get_label_map semantics

Example fix

# before
ds = CityscapesDataset(...,
  metainfo=dict(classes=['road', 'sidewalk', 'car'],
                palette=[[128, 64, 128]]))  # 1 color for 3 classes
# after
ds = CityscapesDataset(...,
  metainfo=dict(classes=['road', 'sidewalk', 'car'],
                palette=[[128, 64, 128], [244, 35, 232], [0, 0, 142]]))
# or simply:
ds = CityscapesDataset(..., metainfo=dict(classes=[...], palette='random'))

Defensive patterns

Strategy: validation

Validate before calling

n_classes = len(metainfo.get('classes', DatasetClass.METAINFO['classes']))
pal = metainfo.get('palette')
if isinstance(pal, list):
    assert len(pal) >= n_classes, f'palette has {len(pal)} colors for {n_classes} classes'

Type guard

def is_palette_valid(palette, n_classes) -> bool:
    return not isinstance(palette, list) or len(palette) >= n_classes

Try / catch

try:
    ds = DatasetClass(..., metainfo=metainfo)
except ValueError as e:
    if 'palette does not match classes' in str(e):
        metainfo['palette'] = 'random'
        ds = DatasetClass(..., metainfo=metainfo)
    else:
        raise

Prevention

Prefer palette='random' unless brand colors are required
Keep palette and classes lists in one metainfo dict so they are edited together, with a paired length test
When subsetting classes, regenerate the palette slice in the same commit

When it happens

Trigger: Passing metainfo=dict(palette=[...]) with fewer color entries than the effective number of classes, e.g., a 3-color palette against a subset of 19 Cityscapes classes, or using a palette sized for the original dataset while classes were extended.

Common situations: Custom visualization configs copied between datasets with different class counts; palette lists with off-by-one lengths; using single-tuple palettes where a list-of-tuples is required.

Related errors

AI-assisted analysis of open-mmlab/mmdetection@cfd5d3a985 (2026-08-27). Data as JSON: /api/errors/8023f443d3954ed4. Report an issue: GitHub.