open-mmlab/mmdetection · error · TypeError

boxes should be Tensor, ndarray, or Sequence, but got {type(

Error message

boxes should be Tensor, ndarray, or Sequence, but got {type(data)}

What it means

BaseBoxes.__init__ (HorizontalBoxes, etc.) only accepts np.ndarray, torch.Tensor, or Python Sequence data. Any other type (int, dict, None, PIL object) raises TypeError before tensor conversion.

Source

Thrown at mmdet/structures/bbox/base_boxes.py:64

        dtype (torch.dtype, Optional): data type of boxes. Defaults to None.
        device (str or torch.device, Optional): device of boxes.
            Default to None.
        clone (bool): Whether clone ``boxes`` or not. Defaults to True.
    """

    # Used to verify the last dimension length
    # Should override it in subclass.
    box_dim: int = 0

    def __init__(self,
                 data: Union[Tensor, np.ndarray, Sequence],
                 dtype: Optional[torch.dtype] = None,
                 device: Optional[DeviceType] = None,
                 clone: bool = True) -> None:
        if isinstance(data, (np.ndarray, Tensor, Sequence)):
            data = torch.as_tensor(data)
        else:
            raise TypeError('boxes should be Tensor, ndarray, or Sequence, ',
                            f'but got {type(data)}')

        if device is not None or dtype is not None:
            data = data.to(dtype=dtype, device=device)
        # Clone the data to avoid potential bugs
        if clone:
            data = data.clone()
        # handle the empty input like []
        if data.numel() == 0:
            data = data.reshape((-1, self.box_dim))

        assert data.dim() >= 2 and data.size(-1) == self.box_dim, \
            ('The boxes dimension must >= 2 and the length of the last '
             f'dimension must be {self.box_dim}, but got boxes with '
             f'shape {data.shape}.')
        self.tensor = data

    def convert_to(self, dst_type: Union[str, type]) -> 'BaseBoxes':

View on GitHub (pinned to cfd5d3a985)

Solutions

1. Ensure data is a Tensor, ndarray, or list/tuple of numbers before wrapping
2. Guard empty annotations: use HorizontalBoxes(torch.zeros(0,4), ...) instead of None
3. Extract the right field from data samples (e.g. results.gt_bboxes.tensor) before re-wrapping

Example fix

# before
boxes = HorizontalBoxes(None) if len(anns) == 0 else ...
# after
import torch
boxes = HorizontalBoxes(torch.zeros(0, 4)) if len(anns) == 0 else ...

import numpy as np, torch
assert isinstance(data, (np.ndarray, torch.Tensor, list, tuple)), type(data)

def is_valid_boxes_data(d) -> bool:
    import numpy as np, torch
    return isinstance(d, (np.ndarray, torch.Tensor, list, tuple))

try:
    boxes = HorizontalBoxes(data)
except TypeError:
    boxes = HorizontalBoxes(torch.zeros(0, 4))

When it happens

Trigger: Constructing HorizontalBoxes with a scalar, None, a generator, or an uninitialized data path result; also passing data that has already been wrapped (e.g. a BaseBoxes instance is not a Sequence).

Common situations: Empty/missing annotations from a dataset sample passed into loss computation, or glue code converting raw dataset dicts to HorizontalBoxes without extracting the array first.

Related errors

• Unsupported {type(mask)} data type
• Only supports dict or list or Tensor, but get {type(results)
• box_list should not be a empty list.
• LoadImageFromFile is not found in the test pipeline
• Visualization needs the "visualizer" termdefined in the conf

AI-assisted analysis of open-mmlab/mmdetection@cfd5d3a985 (2026-08-27). Data as JSON: /api/errors/c76377b83fddc13b. Report an issue: GitHub.