What causes "box_list should not be a empty list." in open-mmlab/mmdetection?

Guard with 'if not box_list: return' or create an empty box instance: cls(torch.zeros(0, 4))

open-mmlab/mmdetection · error · ValueError

box_list should not be a empty list.

Error message

box_list should not be a empty list.

What it means

BaseBoxes.cat concatenates box instances along an existing dim; an empty input list has no boxes to concatenate and no way to infer the box type/shape, so it raises ValueError immediately.

Source

Thrown at mmdet/structures/bbox/base_boxes.py:335

        assert dim != -1 and dim != self.tensor.dim()
        return type(self)(self.tensor.unsqueeze(dim), clone=False)

    @classmethod
    def cat(cls: Type[T], box_list: Sequence[T], dim: int = 0) -> T:
        """Cancatenates a box instance list into one single box instance.
        Similar to ``torch.cat``.

        Args:
            box_list (Sequence[T]): A sequence of box instances.
            dim (int): The dimension over which the box are concatenated.
                Defaults to 0.

        Returns:
            T: Concatenated box instance.
        """
        assert isinstance(box_list, Sequence)
        if len(box_list) == 0:
            raise ValueError('box_list should not be a empty list.')

        assert dim != -1 and dim != box_list[0].dim() - 1
        assert all(isinstance(boxes, cls) for boxes in box_list)

        th_box_list = [boxes.tensor for boxes in box_list]
        return cls(torch.cat(th_box_list, dim=dim), clone=False)

    @classmethod
    def stack(cls: Type[T], box_list: Sequence[T], dim: int = 0) -> T:
        """Concatenates a sequence of tensors along a new dimension. Similar to
        ``torch.stack``.

        Args:
            box_list (Sequence[T]): A sequence of box instances.
            dim (int): Dimension to insert. Defaults to 0.

        Returns:
            T: Concatenated box instance.

View on GitHub (pinned to cfd5d3a985)

Solutions

Guard with 'if not box_list: return' or create an empty box instance: cls(torch.zeros(0, 4))
Check upstream filters that may legitimately produce zero detections before calling cat
Use the dataloader's default collate which handles empty samples

Example fix

# before
all_boxes = HorizontalBoxes.cat(box_list)
# after
all_boxes = (HorizontalBoxes.cat(box_list) if box_list
             else HorizontalBoxes(torch.zeros(0, 4)))

Defensive patterns

Strategy: validation

Validate before calling

assert isinstance(box_list, (list, tuple)) and len(box_list) > 0

Type guard

def can_cat_boxes(bl) -> bool:
    return hasattr(bl, '__len__') and len(bl) > 0

Try / catch

from mmdet.structures.bbox import HorizontalBoxes
import torch
merged = HorizontalBoxes.cat(box_list) if box_list else HorizontalBoxes(torch.zeros(0, 4))

Prevention

Guard aggregation loops that can produce zero items
Return typed empty boxes instead of calling cat on []

When it happens

Trigger: Calling HorizontalBoxes.cat([]) (or via batch collate of zero samples), e.g. looping over per-image boxes in a loop that may execute zero times.

Common situations: Data loaders yielding an empty batch, filter loops removing all boxes, or aggregating detections across an empty frame list in video inference.

Related errors

AI-assisted analysis of open-mmlab/mmdetection@cfd5d3a985 (2026-08-27). Data as JSON: /api/errors/6edb6e6d4c063af4. Report an issue: GitHub.