CanDoYa
All tools
Saved toolsYour favorites

Add to favorites

Saved in this browser only.
Private browser utility / Social & Creator

Free AI Object Detection

Runs entirely in your browser - no upload, no sign-up.

Live workspaceLocal processing

Choose a photo to detect and locate everyday objects.

Share this tool
object detection / browser utility
01 / Overview

What is object detection?

Object detection identifies known object categories in an image and estimates a bounding box around each match. This tool runs a compact YOLOS model inside your browser, lists confidence scores, and exports an annotated PNG. Your image stays on your device while the public AI runtime and model download separately.

02

How to use

  1. 01
    Choose an image

    Drop in one JPEG, PNG, or WebP up to 20 MB and 16 megapixels, or load the sample scene to explore the interface.

  2. 02
    Run detection

    Click Detect objects. The worker downloads the public model on the first run, uses WebGPU when available, and falls back to CPU processing when needed.

  3. 03
    Review and export

    Adjust minimum confidence without rerunning the model, inspect every visible box and score, then copy the report or download the annotated PNG.

03

Who it's for

  • Computer vision students who want to inspect labels, boxes, and threshold effects without installing Python or a model runtime.
  • Dataset reviewers who need a quick first pass over common COCO categories before manual annotation.
  • Creators and developers checking whether a scene contains recognizable people, vehicles, animals, furniture, or electronics.
  • Private image workflows where sending an unreleased photo or screenshot to a cloud detection API is not appropriate.

This detector uses the Apache-2.0 YOLOS-tiny model through Transformers.js 4.2.0 in a dedicated browser worker. It prefers roughly 13 MB of FP16 ONNX weights on WebGPU and automatically retries with roughly 10 MB of q8 weights through WebAssembly on the CPU. Runtime files download separately, and browser caching can make later runs faster.

YOLOS-tiny was fine-tuned on COCO 2017 and predicts 80 everyday categories such as person, bicycle, car, dog, cat, chair, bottle, book, and laptop. It is not an open-vocabulary identifier. A lower confidence threshold reveals more tentative boxes but also more false positives; a higher threshold shows fewer, stronger matches.

FAQ

Is my image uploaded to a server?

No. The selected image is decoded, passed to a worker, analyzed, annotated, and encoded inside your browser. The worker downloads the public Transformers.js runtime and YOLOS-tiny model from external hosts, but those requests do not contain your image. CanDoYa does not receive or store the source or result.

Is this object detector free?

Yes. There is no account, payment, credit limit, or watermark. Your device supplies the computing power. The first run still downloads public runtime and model files, so it uses your internet connection and may matter on a metered plan.

What image limits apply?

The tool accepts one JPEG, PNG, or WebP image up to 20 MB and 16 megapixels, with each side at least 32 pixels. These limits reduce large browser memory spikes. The annotated PNG keeps the original pixel dimensions and can be larger than the source because PNG uses lossless compression.

What objects can the AI detect?

YOLOS-tiny predicts the 80 categories used by the COCO dataset, including people, common vehicles, animals, sports gear, food, furniture, appliances, and electronics. It cannot name arbitrary products, plant species, landmarks, text, or concepts outside that fixed label set.

How accurate is YOLOS-tiny object detection?

The compact model reports 28.7 average precision on COCO 2017 validation. Clear, well-lit images with larger familiar subjects tend to work better. Tiny objects, unusual angles, heavy overlap, blur, poor lighting, and categories outside COCO can be missed or mislabeled, so always review the boxes.

What does the confidence score mean?

Confidence ranks how strongly the model supports a label and box for this image. It is not a guarantee or a calibrated probability that the object is present. Lowering the threshold shows more uncertain matches; raising it hides weaker boxes and can reduce visible false positives.

Does object detection work without WebGPU?

Yes, if the browser supports WebAssembly and module workers. The worker checks for a usable WebGPU adapter first. If none exists or WebGPU inference fails, it retries with q8 model weights on the CPU through WebAssembly. The fallback stays on-device but can take longer.

Can I use the object detector offline?

A later run may work without another download after the browser caches the runtime and model, but offline availability is not guaranteed. Browsers can evict cached files, storage cleanup can remove them, and private browsing often discards them when the session ends.