Object Detector
Automatically detect objects in a photo — people, vehicles, animals, and common everyday items — using a detection model that runs entirely on your device via WebAssembly. Adjust the confidence threshold instantly without re-running the model, then export straight into YOLO, COCO, Pascal VOC, or VIA CSV — ready to open in the Image Annotator to fix anything the model missed.
About this tool
This tool automatically detects objects in a photo — people, vehicles, animals, and common everyday items — using a real object detection model that runs entirely on your device via WebAssembly, then exports the result directly as YOLO, COCO, Pascal VOC, or VIA CSV annotations. It exists to skip the slowest part of building an object detection dataset: drawing every single box by hand. Instead, upload an image, let the on-device model suggest boxes, adjust the confidence threshold until the results look right, and export — then open the result in the Image Annotator to fix anything the model missed or mislabeled, exactly the same workflow as verifying and correcting any other annotation source. The confidence slider re-filters already-computed detections instantly; it doesn't need to re-run the model each time you move it. Nothing about your photo is ever uploaded — the model itself (~7MB) downloads once from Tuliq and every image after that processes completely offline.
Supported formats
- YOLO (.txt) — One .txt file plus a classes.txt — normalized 0–1 coordinates, exactly what bbox-converter and image-annotator also produce.
- COCO (.json) — A single JSON file with your detections as COCO-format annotations, absolute pixel coordinates.
- Pascal VOC (.xml) — One XML file with your image's detections as <object> blocks.
- VIA CSV (.csv) — One CSV row per detected object, in VIA's native format.
What you can do
- Automatically detect objects in a photo and draw bounding boxes
- Runs a real detection model on your own device, not a server
- Export straight into YOLO, COCO, Pascal VOC, or VIA CSV
- Your image is never uploaded — detection runs entirely in your browser
How to use
Step 1
Upload a photo
The model recognizes about 90 everyday object categories — people, vehicles, animals, furniture, and more.
Step 2
Review the detections
Adjust the confidence slider instantly — no need to re-run the model to filter out weak guesses.
Step 3
Export the annotations
Pick a format and download — ready to open directly in the Image Annotator or Annotation Visualizer.
Example
A detected object comes back from the model as a class name, a confidence score, and a pixel-space bounding box:
Raw detection
{ class: "car", score: 0.92, box: [140, 210, 320, 180] }That's a car detected with 92% confidence, at pixel (140, 210), 320px wide and 180px tall — which this tool then hands straight to the same annotation format serializers bbox-converter uses, so the exported file is indistinguishable from one you drew by hand.
How your data is handled
This tool is labeled 🔵 Local AI: your photo is read directly using your browser's File API and detection runs entirely on your device via a local WebAssembly model. It's never uploaded anywhere. The one network request this tool makes is downloading the model file (~7MB) once, which is cached by your browser afterward — every photo you process after that runs completely offline.
Frequently asked questions
Is my photo uploaded anywhere?+
No — it's read directly in your browser and detection runs on your device via a local WebAssembly model. It never leaves your device. The only network request is a one-time model download (~7MB), cached by your browser afterward.
What objects can it detect?+
The model recognizes about 90 everyday categories — people, vehicles, animals, furniture, food, and common household items. It's a general-purpose detector, not trained on anything domain-specific, so unusual or specialized objects won't be recognized.
Why did moving the confidence slider not take a moment to reprocess?+
The model only runs once per photo. Every detection above a low baseline confidence is kept from that single run; the slider just re-filters that same result instantly, entirely in your browser, without needing to re-run the model.
Can I fix a wrong or missing detection before exporting?+
Not directly in this tool — export what the model found, then open the result in the Image Annotator to add, remove, or relabel boxes before using the dataset.
Why does it say two files downloaded for a YOLO export?+
YOLO's .txt format has no way to store class names, only numeric indices — the classes.txt companion (listing names in index order) is the second file, same as every other Tuliq tool that exports YOLO.