opendatalab/MinerU · error · ValueError

Input image ({w}, {h}) smaller than the target size ({cw}, {

Error message

Input image ({w}, {h}) smaller than the target size ({cw}, {ch}).

What it means

The table-orientation classifier resizes the short edge to 256 then center-crops 224x224; after that resize both dimensions should be >= 224, so this guard fires only for degenerate inputs (near-zero-size images where scaling cannot produce a valid crop, or rounding on extreme aspect ratios). It protects the downstream slice from producing an empty tensor.

Source

Thrown at mineru/model/table/cls/paddle_table_cls.py:43

        self.mean = [0.485, 0.456, 0.406]
        self.labels = [AtomicModel.WiredTable, AtomicModel.WirelessTable]

    def preprocess(self, input_img):
        # 放大图片，使其最短边长为256
        h, w = input_img.shape[:2]
        scale = 256 / min(h, w)
        h_resize = round(h * scale)
        w_resize = round(w * scale)
        img = cv2.resize(input_img, (w_resize, h_resize), interpolation=1)
        # 调整为224*224的正方形
        h, w = img.shape[:2]
        cw, ch = 224, 224
        x1 = max(0, (w - cw) // 2)
        y1 = max(0, (h - ch) // 2)
        x2 = min(w, x1 + cw)
        y2 = min(h, y1 + ch)
        if w < cw or h < ch:
            raise ValueError(
                f"Input image ({w}, {h}) smaller than the target size ({cw}, {ch})."
            )
        img = img[y1:y2, x1:x2, ...]
        # 正则化
        split_im = list(cv2.split(img))
        std = [0.229, 0.224, 0.225]
        scale = 0.00392156862745098
        mean = [0.485, 0.456, 0.406]
        alpha = [scale / std[i] for i in range(len(std))]
        beta = [-mean[i] / std[i] for i in range(len(std))]
        for c in range(img.shape[2]):
            split_im[c] = split_im[c].astype(np.float32)
            split_im[c] *= alpha[c]
            split_im[c] += beta[c]
        img = cv2.merge(split_im)
        # 5. 转换为 CHW 格式
        img = img.transpose((2, 0, 1))
        imgs = [img]

View on GitHub (pinned to 4fe4bde114)

Solutions

1. Validate the crop size upstream: skip regions with w < 10 or h < 10 before classification.
2. Check the layout detection boxes for zero/negative area and clamp them.
3. If tiny tables must be classified, upscale with a stronger interpolation before this transform.

Example fix

# before
cls_result = table_cls.predict(crop)  # crop may be 5x5

# after
if min(crop.shape[:2]) < 24:
    continue  # skip degenerate crop
cls_result = table_cls.predict(crop)

h, w = crop.shape[:2]
if min(h, w) < 24:
    skip = True  # degenerate crop; do not classify

def is_classifiable_crop(img: np.ndarray) -> bool:
    return img.ndim >= 2 and min(img.shape[:2]) >= 24

try:
    res = table_cls(img)
except ValueError as e:
    if 'smaller than the target size' in str(e):
        return default_orientation()  # skip/assume no rotation
    raise

When it happens

Trigger: Calling the cls transform with an image whose min(h, w) is 0 or 1 (blank crop, failed region extraction), producing a resized dimension below 224 after integer rounding.

Common situations: Empty table-region crops from an upstream layout detector with bad boxes (negative or zero-area), fully transparent/blank page regions rendered at tiny size.

Related errors

• The channel({channel}) of the img is not in [1, 2, 3, 4]
• backend: {backend} is not supported for resize.Supported bac
• Unsupported image shape for UnimerSwinImageProcessor: {image
• The img type {type(img)} does not in {InputType.__args__}
• cannot identify image file {img}

AI-assisted analysis of opendatalab/MinerU@4fe4bde114 (2026-08-14). Data as JSON: /api/errors/424c78a98184420d. Report an issue: GitHub.