Drawing every bounding box by hand is the slowest part of building an object detection dataset. The Object Detector skips the blank-page problem: upload an image, let an on-device model suggest boxes, then treat those suggestions as a draft to correct rather than a dataset to draw from scratch.
The workflow
- Upload an image to the Object Detector. The model runs once, entirely on your device, and returns every object it found above a low confidence baseline.
- Adjust the confidence slider. This re-filters the same result instantly — no need to re-run the model to hide weak, low-confidence guesses. Raise it until what's left looks right; lower it if you suspect the model missed something borderline.
- Export. Pick YOLO, COCO, Pascal VOC, or VIA CSV — the same four formats and the same serializers the Bounding Box Converter uses, so the file this produces is indistinguishable from one drawn by hand.
- Open the export in the Image Annotator. This is the step that actually matters: treat the auto-detected boxes as a first draft. Add anything the model missed, remove false positives, fix a wrong class, or adjust a box that's slightly off. A model detection is a starting point, not a finished annotation.
- Verify the final result in the Annotation Visualizer before trusting the dataset for training — the same verification step that matters for any annotation source, covered in how to visually verify annotations before training.
Why this model, specifically
The Object Detector runs EfficientDet-Lite0, a general-purpose detector trained on roughly 90 everyday object categories. It's a deliberate choice worth being upfront about: it's not the most accurate detector available. Newer models like YOLOv8 outperform it on raw benchmark accuracy — but YOLOv8's default license (AGPL-3.0) carries real copyleft obligations that don't fit a tool meant to be freely embedded in a public site. EfficientDet-Lite0 is Apache 2.0 — permissively licensed, purpose-built for on-device deployment, and safe to actually ship. Between a more accurate model we couldn't responsibly distribute and a genuinely good one we could, this tool takes the second option.
In practice, this means: expect solid results on common objects (people, vehicles, animals, everyday items), and expect to correct rather than blindly trust the output — which is true of any auto-detection tool, and exactly why step 4 above isn't optional.
When this is and isn't the right starting point
Auto-detection saves the most time when your dataset is full of the kind of objects the model already recognizes — people, cars, common household items. It saves little to no time on a specialized or domain-specific dataset (industrial defects, medical imagery, a custom product category) the model was never trained to recognize — for that, drawing from scratch in the Image Annotator is still the right starting point, not a detour.