Dataset Health Check
Inspect image validity, dimensions, missing labels, and orphan annotation files.
Read the guide and worked example ↓
No settings are required. The tool will create a safe report or output without changing your original file.
Recent Results
The download appears directly below its progress bar when processing finishes.
How to use Dataset Health Check
Use this check before annotation, conversion, or training to find unreadable images and mismatched file names. It checks the structure of an image dataset; it does not decide whether an annotation describes the correct object.
Prepare, run, and review
- Prepare a ZIP containing supported images and their annotation files. Give each image a unique base name across the entire archive. Matching currently uses the base name, not the full directory path.
- For example, pair images/frame_001.jpg with labels/frame_001.txt. Remove unrelated JSON, XML, and text documents from the test archive: these extensions are treated as possible label files.
- Upload the ZIP and run the check. The account notice shows the applicable quota and upload limit; processing does not edit the original ZIP.
- Download dataset_health_report.zip. Read the summary JSON, then inspect the per-image CSV when present to locate individual corrupt or unmatched images.
Download a small, synthetic example dataset. It contains four labeled rectangle images and a matching copy of the first image for duplicate-check practice. It is for learning the file workflow, not for training a useful detector.
How to read a small report
{"images": 3, "labels": 3, "corrupt_images": 0,
"images_without_labels": 1, "orphan_labels": 1}This illustrative report has equal image and label totals, but the names do not all match. One image has no matching label, and one label has no matching image. Equal totals alone do not prove a complete dataset. Compare the base names before adding or deleting anything.
Limits and common mistakes
- A missing label may be intentional for an unlabeled image. Apply your dataset's labeling policy before treating it as an error.
- A matching label name does not prove that its contents are valid. Use Annotation Quality Checker for supported YOLO box rows, then visually review the boxes.
- Corrupt images should be replaced from a trustworthy original where possible. Renaming an unreadable file does not repair its contents.
- Mixed image dimensions may be legitimate. Decide whether resizing is necessary and transform annotations consistently if you do resize.
The image and label totals are equal. Why is one image still reported as missing a label?
The check pairs base names. A different label can be orphaned while the image lacks its own matching label, so the totals can still be equal.
Continue learning
ToolsInVision guide · Updated September 26, 2026 · Editorial approach · Report an issue