What causes "Image '{}' format not JPEG" in WZMIAOMIAO/deep-learning-for-image-processing?

Re-encode the offending image(s) to JPEG with PIL: Image.open(p).convert('RGB').save(p, 'JPEG').

WZMIAOMIAO/deep-learning-for-image-processing · error · ValueError

Image '{}' format not JPEG

Error message

Image '{}' format not JPEG

What it means

VOCDataSet.__getitem__ opens each image referenced by a VOC XML annotation and requires its PIL image.format to be 'JPEG'. The dataset assumes VOC-style JPEG images; other formats (PNG, BMP, GIF) cannot pass the standard VOC torchvision-style pipeline (and later transforms may fail on non-RGB data). It throws a ValueError naming the offending image path.

Source

Thrown at pytorch_object_detection/retinaNet/my_dataset.py:56

        with open(json_file, 'r') as f:
            self.class_dict = json.load(f)

        self.transforms = transforms

    def __len__(self):
        return len(self.xml_list)

    def __getitem__(self, idx):
        # read xml
        xml_path = self.xml_list[idx]
        with open(xml_path) as fid:
            xml_str = fid.read()
        xml = etree.fromstring(xml_str)
        data = self.parse_xml_to_dict(xml)["annotation"]
        img_path = os.path.join(self.img_root, data["filename"])
        image = Image.open(img_path)
        if image.format != "JPEG":
            raise ValueError("Image '{}' format not JPEG".format(img_path))

        boxes = []
        labels = []
        iscrowd = []
        assert "object" in data, "{} lack of object information.".format(xml_path)
        for obj in data["object"]:
            xmin = float(obj["bndbox"]["xmin"])
            xmax = float(obj["bndbox"]["xmax"])
            ymin = float(obj["bndbox"]["ymin"])
            ymax = float(obj["bndbox"]["ymax"])
            boxes.append([xmin, ymin, xmax, ymax])
            labels.append(self.class_dict[obj["name"]])
            if "difficult" in obj:
                iscrowd.append(int(obj["difficult"]))
            else:
                iscrowd.append(0)

        # convert everything into a torch.Tensor

View on GitHub (pinned to 1ec3fe6f37)

Solutions

Re-encode the offending image(s) to JPEG with PIL: Image.open(p).convert('RGB').save(p, 'JPEG').
Batch-convert all non-JPEG files in the dataset directory before training (script over the img_root checking image.format).
If PNG support is needed, relax the check and convert on load instead of raising: image = Image.open(img_path).convert('RGB').
Fix filename/annotation mismatches: ensure data['filename'] in the XML points at the real JPEG file.

Example fix

// before
image = Image.open(img_path)
if image.format != "JPEG":
    raise ValueError("Image '{}' format not JPEG".format(img_path))
// after
image = Image.open(img_path)
if image.format != "JPEG":
    image = image.convert("RGB")  # tolerate non-JPEG input

Defensive patterns

Strategy: validation

Validate before calling

from PIL import Image
for name in os.listdir(img_root):
    with Image.open(os.path.join(img_root, name)) as im:
        if im.format != "JPEG":
            print("non-JPEG:", name, im.format)

Type guard

def is_jpeg(path: str) -> bool:
    from PIL import Image
    try:
        with Image.open(path) as im:
            return im.format == "JPEG"
    except Exception:
        return False

Try / catch

try:
    sample = dataset[i]
except ValueError as e:
    if "format not JPEG" in str(e):
        img = Image.open(extract_path(e)).convert("RGB")
        img.save(extract_path(e), "JPEG")
    else:
        raise

Prevention

Batch-verify image formats right after downloading a dataset.
Never rely on file extensions; check actual image.format.
Re-encode through PIL/convert('RGB') during dataset preparation.
Keep annotation filenames in sync with actual image files.

When it happens

Trigger: Iterating the dataset (DataLoader over VOCDataSet) when the file at os.path.join(self.img_root, data['filename']) is not a JPEG: e.g. a PNG renamed to .jpg, a grayscale/CMYK image still saved as PNG, or a mixed-format directory indexed by train.txt.

Common situations: Downloading images from the web where extensions lie; converting a dataset with scripts that copy files without re-encoding; VOC2007/VOC2012 datasets containing some PNG segmentation images mixed into JPEGImages; re-labeling data with annotation tools that save PNG.

Related errors

AI-assisted analysis of WZMIAOMIAO/deep-learning-for-image-processing@1ec3fe6f37 (2026-08-30). Data as JSON: /api/errors/3ceb7bb0774b8e62. Report an issue: GitHub.