COCO Dataset and Annotation Format

The COCO project was created to study common objects in realistic context, with annotations supporting several computer-vision tasks. People often say “COCO format” to mean a JSON structure inspired by that project, even when their images are completely different. That convention can make tools interoperate, but it says little about image quality, label accuracy or suitability for a deployment setting. A model trained on household scenes may not work on industrial components just because both datasets use similar JSON keys. For object detection, an annotation refers to an image and a category, typically through image_id and category_id fields. A bounding box is commonly stored as x and y of its upper-left corner plus width and height, rather than the coordinates of the opposite corner. Each image can contain many object instances and each instance has its own annotation record. Segmentation tasks add information about object shape, often polygons or encoded masks. Captions, keypoints and panoptic tasks use their own associated conventions and evaluation routines; do not treat every COCO file as interchangeable. Reliable conversion requires checking identifiers, dimensions and coordinate systems. A box written in normalized 0-to-1 coordinates but interpreted as pixels will be almost invisible. A category mapping that changes integer IDs between train and test can silently corrupt results. Visualize random annotations over source images, including small and crowded objects, and count missing references, impossible boxes and empty masks. Keep train and test images separated by scene or source where necessary to prevent near-duplicate leakage. The original COCO paper and official dataset tools establish the reference context and APIs. When repackaging a new dataset in the format, document exactly which version or task schema a tool expects. Evaluate on examples resembling real use, because annotation compatibility is a software convenience, not a guarantee of model generalization or fairness.