Skip to main content
Preparing workspace

Live Object Detector

Free live object detection in your browser: COCO-SSD boxes, labels and confidence from your webcam or a photo, with FPS, counts and screenshots. No upload.

Still loading. If the tool doesn’t appear, reload this page.

About this tool

Live Object Detector

Detect objects live from your webcam or an uploaded image with COCO-SSD in your browser: boxes, labels, confidence, FPS, per-class counts and annotated screenshots.

Live Object Detector runs the COCO-SSD model on TensorFlow.js inside your browser tab to find everyday objects in a live webcam feed or an uploaded image. Each detection gets a coloured bounding box, a class label and a confidence percentage, drawn over the picture and kept aligned as the window resizes. A dashboard beside the video shows frames per second, per-frame inference time, how many objects are in view and a per-class table with the peak count and best confidence for the session. You can pause on a frame, move a confidence slider to see boxes appear and disappear, save annotated PNG screenshots, and keep a local history of sessions and images. The model detects the 80 COCO classes only, such as person, car, dog, cup, laptop and cell phone, and the video never leaves your device.

How to use Live Object Detector

  1. 1

    Pick Webcam and press Start camera, then allow camera access when the browser asks. The model downloads once (about 4 MB) and detection starts by itself.

  2. 2

    Point the camera at everyday objects and watch the labelled boxes, the FPS tile, the in-view count and the Detected objects table update live.

  3. 3

    Move the Confidence threshold slider to show more or fewer boxes, and use Pause to freeze a frame, then Resume or Stop when you are done.

  4. 4

    Press Capture screenshot to save an annotated PNG, or choose the Image tab and upload a photo to detect objects in it.

Why use Live Object Detector

  • Runs entirely in your browser

    TensorFlow.js executes the model on your GPU through WebGL (or the CPU as a fallback). There is no server, account, API key or cloud GPU, and your camera frames are never uploaded.

  • Honest live numbers

    FPS is the real count of completed detections per second, and the inference time is measured per frame. Only one detection runs at a time, so a slow machine lowers the frame rate instead of building a backlog.

  • Threshold without re-running

    The model reports everything above 10% and the slider filters the boxes afterwards, so changing the confidence updates the picture and the counts immediately.

  • Boxes that stay aligned

    The overlay matches the video or image exactly at any size, including mirrored selfie view, and the saved screenshot draws the same boxes at full resolution.

Common use cases

  • You want to see what an off-the-shelf computer-vision model recognises in your room, desk or street before deciding whether it is good enough for a project.
  • You are teaching or demonstrating object detection and need a live example with visible boxes, confidence scores and a threshold slider that responds as you drag it.
  • You have a photo of a shelf, a car park or a crowd and want boxed, labelled detections with per-class counts that you can save as an annotated image.

Frequently asked questions

What objects can the Live Object Detector recognise?

It recognises the 80 object classes of the COCO dataset: people, common animals, vehicles, street items, food, kitchen items, furniture and electronics such as laptop, tv, cell phone and keyboard. Anything outside that list is not detected, and a similar object may be given the nearest class.

Does it recognise faces, text or specific brands?

No. COCO-SSD only knows 80 general object classes. It does not identify people, read text or logos, or detect specialised items such as tools or machine parts.

Is my webcam video uploaded anywhere?

No. Frames are read from the camera in your browser and processed on your device by TensorFlow.js. Only the model weights are downloaded, once, from a public Google storage address. Stopping the camera or leaving the page releases the camera.

Why is the frame rate low?

Detection runs the neural network on every frame, so speed depends on your GPU and browser. If WebGL is unavailable the tool falls back to the CPU, which is slower. Closing other tabs and using a lower-resolution camera helps.

What does the confidence threshold do?

It hides detections the model is less sure about. At 50% only objects scored 50% or higher are boxed and counted. Lower values find more objects but add false ones; higher values are stricter.

How do I save an annotated image?

Press Capture screenshot (or the C key) while the camera is running or after loading an image. The PNG appears under Annotated screenshots with a download button and contains the picture with every visible box and label.

Why did the camera not start?

The browser may have blocked camera access, another app may be using the camera, or the page may not be secure. Allow the camera for this site from the address bar, close other apps that use it, and press Start camera again. You can still upload an image.

Can I use my phone?

Yes. On a phone the layout stacks vertically and the camera list shows your front and rear cameras. A rear camera is shown unmirrored; use the Mirror button to flip a front camera.

Which image files can I upload?

Any image the browser can open, such as JPG, PNG, WebP or GIF, up to 25 MB. The image is analysed locally and is not uploaded.

What does the history keep?

Stopped camera sessions, analysed images and screenshots are listed with their counts. Only the text summary is stored in this browser, never the pictures, and Clear history removes it.