fix: return three values when a gif cannot be decoded - #18346

🔗 Related Issue

None — found while reading ppocr/utils/utility.py.

📖 Description

check_and_read returns a 3-tuple from every branch except one. When cv2.VideoCapture cannot decode a gif, it returns (None, False):

https://github.com/PaddlePaddle/PaddleOCR/blob/main/ppocr/utils/utility.py#L123-L126

Every caller unpacks three values. From tools/infer/predict_system.py:212:

img, flag_gif, flag_pdf = check_and_read(image_file)

predict_rec.py:892, predict_cls.py:146, predict_e2e.py:161 and predict_sr.py:136 do the same with img, flag, _ =. So a corrupted gif raises in the caller instead of being skipped:

ValueError: not enough values to unpack (expected 3, got 2)

That defeats the branch's own intent — it logs a message and returns a sentinel precisely so the caller can move to the next file.

There is a second problem in the same three lines. The log call passes an unformatted {}:

logger.info("Cannot read {}. This gif image maybe corrupted.")

so the output drops the path that was the only thing it existed to report:

[    INFO] utility.py:125 - Cannot read {}. This gif image maybe corrupted.

Reproduction

from ppocr.utils.utility import check_and_read

with open("broken.gif", "wb") as f:      # .gif name, undecodable contents
    f.write(b"GIF89a" + b"\x00\xff garbage")

img, flag_gif, flag_pdf = check_and_read("broken.gif")

Before this change, on paddlepaddle 3.3.1 / Python 3.12:

[    INFO] utility.py:125 - Cannot read {}. This gif image maybe corrupted.
ValueError: not enough values to unpack (expected 3, got 2)

After:

[    INFO] utility.py:125 - Cannot read /tmp/.../broken.gif. This gif image maybe corrupted.
img=None flag_gif=False flag_pdf=False

(None, False, False) matches what the function already returns for a path it does not handle (utility.py:152), so callers treat the corrupted gif the same way they treat any non-gif, non-pdf input.

🧪 Tests

Added tests/ppocr/test_check_and_read.py: one test per bug, plus a check that non-gif and non-pdf paths keep their current return value. That third test does not depend on the fix and stays green on either side of it.

The first two are load-bearing — restoring the old two-value return and the unformatted log message turns exactly those red:

FAILED tests/ppocr/test_check_and_read.py::test_corrupted_gif_returns_three_values
FAILED tests/ppocr/test_check_and_read.py::test_corrupted_gif_logs_the_path
2 failed, 4 passed

With the fix in place:

tests/ppocr/test_check_and_read.py                     6 passed
tests/ (excluding tests/ppocr/test_iaa_augment.py)     126 passed, 89 errors
black --check                                          2 files unchanged

The 89 errors are pre-existing on this machine and unrelated: they are all in tests/pipelines/, which downloads model weights this box cannot reach, and test_iaa_augment.py needs albumentations. I confirmed via git stash that tests/pipelines/test_table_recognition_v2.py produces the identical errors on a clean checkout.

✅ Checklist

• Ran black on changed files
• Added tests that fail without the change
• No functional change to any other branch of the function

AI assistance was used for this change; I reviewed every changed line and ran the commands above locally.