Bounding box annotation is the backbone of most object detection datasets, and small inconsistencies in how boxes are drawn show up later as lost precision. ImageLabeling draws the first boxes for you, lets your labelers fix them with the keyboard, and keeps every team on the same rules for tight boxes, occluded objects and objects cut off by the frame.
Press B, drag from one corner to the opposite corner, pick the class and move on. Corners and edges stay draggable, so a box that is a few pixels off takes one move to fix. Zoom in for small objects, and undo history keeps every change reversible while you work. Each box stores its class, its coordinates and, on Team and above, attributes you define in the schema.
Before a labeler opens an image, the AI model proposes boxes for the classes in your label schema, each with a confidence score. Labelers sort by confidence, accept the clear ones and spend their time on the uncertain ones. Every proposal still passes through a human, and nothing reaches export unreviewed when review is on. Read how this works on the auto labeling page.
A detector learns the edges you draw. If half your team leaves a margin and the other half does not, the model learns a blur of both. Write the rules into the project once and every labeler sees them.
Boxes stay axis aligned, so long diagonal objects pull in a lot of background. For pipes, cracks or rotated parcels, a polygon from the image segmentation tool describes the object better, and the COCO export also stores the enclosing box of every polygon.
An object partly hidden behind another object is occluded. The common convention is to box only the visible part and mark the object as occluded with an attribute. Some teams prefer to box the full estimated extent. Either works, as long as one convention is written down and used across the dataset. Attributes per object are available on Team and above.
An object cut off by the image border is truncated. Draw the box to the image edge, not beyond it, and set a truncated attribute so you can filter these objects at training time. Very small fragments that no person could identify should usually stay unlabeled, and your guidelines should say how small is too small.
Dense groups, such as a pallet of identical boxes or a crowd in the background, are hard to label one by one. Decide per class whether you label every instance or skip objects below a size limit, and record that choice in the project guidelines.
A reviewer sees every submitted image, approves it or sends it back with a comment pinned to the box in question. On Team and above, consensus labeling on a sample of images shows where labelers disagree, which is usually where the rules are unclear. Business adds a QA score per labeler.
Boxes export to COCO JSON as x, y, width and height in pixels, to YOLO txt as normalized center coordinates, to Pascal VOC XML as corner coordinates and to CSV. Pick the format at export time, not at labeling time. See the YOLO labeling tool and COCO annotation tool pages for the exact structure, or compare all tools on the data annotation tools page.
Bounding boxes, polygons, points and image classes are part of every plan, from Starter at 24 USD per month billed yearly. Attributes for occlusion and truncation start on Team. The full limits per plan are on the pricing page.
Draw the visible part and mark the object as occluded, unless your model needs the full extent. The most important thing is that every labeler follows the same rule.
The AI model proposes boxes with a confidence score, and they are a first draft. Your labelers adjust edges with a drag and accept the result, so the final box follows your rules.
Yes. Import existing labels as COCO JSON, YOLO txt, Pascal VOC XML or CSV, then review and correct them in the same editor.
Boxes are axis aligned. For angled objects, use a polygon. It keeps the exact outline, and the COCO export also stores its enclosing box.
Upload one image to the free demo, choose Detect objects and adjust the boxes in the real editor.
1 / 2
Keep your labeled image, unlock every object and export it in COCO, YOLO or Pascal VOC.
By continuing you accept the Terms and the Privacy Policy.