Inspect a COCO dataset
Drop your annotations and images to see every box, mask and keypoint, check the labels and score predictions. All in your browser.
1Annotationsrequired
COCO JSON with images, annotations and categories.
or drop it here
2Imagesoptional
A folder, image files or a .zip. Matched by file name.
or drop them here
3Predictionsoptional
A COCO results list: image_id, category_id, bbox or segmentation, score.
or drop it here
12 drawn tabletop scenes with masks, keypoints, a crowd region, planted label mistakes and a predictions file.
What this reads and how the numbers are computed
- Annotations: the COCO object-detection format. Boxes are
[x, y, width, height]in pixels. Masks can be polygons, uncompressed RLE (crowd regions) or compressed RLE strings. Keypoints are[x, y, v]triples with the skeleton from the category. - Images are matched to
file_nameby full path, then by path ending, then by base name. Without images you still see every annotation on a blank canvas of the right size. - Large files: JSON over 48 MB is read in a streaming pass, one record at a time, so the full COCO train annotations open in a desktop browser. Zip archives are read from their index; each image is unpacked only when shown.
- Metrics follow the standard COCO evaluation: IoU thresholds 0.50 to 0.95 in steps of 0.05, 101 recall points, at most 100 detections per image (20 for keypoints), small/medium/large at 32² and 96² pixels, and crowd regions as ignore. Results were checked against the reference implementation to 1e-14 on test sets for boxes, masks and keypoints.
- Per-class precision and recall use one IoU threshold and one score threshold, with the same greedy, highest-score-first matching.
- Privacy: files are read by your browser and never sent anywhere. Closing the tab forgets them.