facebookresearch/detectron2 · error · SizeMismatchError

Mismatched image shape{}, got {}, expect {}.

Error message

Mismatched image shape{}, got {}, expect {}.

What it means

check_image_size verifies that the actually decoded image's (width, height) matches the 'width'/'height' fields recorded in the dataset dict. A mismatch means the metadata is stale relative to the image files, and detectron2 raises SizeMismatchError rather than silently training on wrong geometry.

Source

Thrown at detectron2/data/detection_utils.py:197

    """
    with PathManager.open(file_name, "rb") as f:
        image = Image.open(f)

        # work around this bug: https://github.com/python-pillow/Pillow/issues/3973
        image = _apply_exif_orientation(image)
        return convert_PIL_to_numpy(image, format)
    raise ValueError(f"Failed to read image at: {file_name}")


def check_image_size(dataset_dict, image):
    """
    Raise an error if the image does not match the size specified in the dict.
    """
    if "width" in dataset_dict or "height" in dataset_dict:
        image_wh = (image.shape[1], image.shape[0])
        expected_wh = (dataset_dict["width"], dataset_dict["height"])
        if not image_wh == expected_wh:
            raise SizeMismatchError(
                "Mismatched image shape{}, got {}, expect {}.".format(
                    (
                        " for image " + dataset_dict["file_name"]
                        if "file_name" in dataset_dict
                        else ""
                    ),
                    image_wh,
                    expected_wh,
                )
                + " Please check the width/height in your annotation."
            )

    # To ensure bbox always remap to original image size
    if "width" not in dataset_dict:
        dataset_dict["width"] = image.shape[1]
    if "height" not in dataset_dict:
        dataset_dict["height"] = image.shape[0]

View on GitHub (pinned to a2f4a8771a)

Solutions

1. Re-generate dataset dicts (re-run the loader/json generation) against the current image files
2. Verify and fix the file_name paths and dimensions: check cv2.imread(f).shape vs record['width']/['height']
3. If images were resized, update width/height in the json accordingly

Example fix

# before
# json says width=1920, height=1080 but image on disk is 1280x720
# after
import cv2
for r in DatasetCatalog.get('mydata_train'):
    h, w = cv2.imread(r['file_name']).shape[:2]
    assert (w, h) == (r['width'], r['height']), r['file_name']

import cv2
for r in dataset_dicts:
    h, w = cv2.imread(r['file_name']).shape[:2]
    if (w, h) != (r.get('width', w), r.get('height', h)):
        print('mismatch:', r['file_name'], (w, h), (r['width'], r['height']))

def dims_match(record, img) -> bool:
    h, w = img.shape[:2]
    return (w, h) == (record['width'], record['height'])

from detectron2.data.detection_utils import SizeMismatchError
try:
    out = mapper(dataset_dict)
except SizeMismatchError as e:
    log.warning(f'skipping stale record: {e}')
    return None

When it happens

Trigger: Reading a dataset whose json/metadata records different dimensions than the images on disk — e.g. images were resized/re-exported after annotation json was generated, or file_name points to a different file than the one annotated.

Common situations: Regenerating/resizing images without refreshing the annotation json; mixing train/val image directories; symlinks or caching (LMDB/SerializeList) serving old images; metadata copied from another split.

Related errors

• Cannot match one checkpoint key to multiple keys in the mode
• Class with @configurable must have a 'from_config' classmeth
• {name} must take 'cfg' as the first argument!
• target of LazyCall must be a callable or defines a callable!
• total_batch_size and single_gpu_batch_size are mutually inco

AI-assisted analysis of facebookresearch/detectron2@a2f4a8771a (2026-08-27). Data as JSON: /api/errors/7881de04f6b0cf21. Report an issue: GitHub.