About Online Object Detector
This online object detector is free, needs no sign-up and runs in your browser, so your images are never uploaded. Choose a photo, paste one, take one with your phone camera or start your webcam, and it draws a box around each object it recognises, labelled with the object name and a score, for example person 91% or dog 87%.
It uses COCO-SSD, a pretrained detection model that recognises the 80 object classes of the COCO dataset: people, vehicles, common animals, furniture, kitchen items, food, electronics and sports gear. You choose between two versions of it, lite (fast) at about 18 MB and accurate at about 67 MB, and the size is shown before anything downloads. It is not YOLO, and it only detects those 80 classes, so a tool or a plant species that is not on the list will not be labelled.
One thing is downloaded. Your images and camera frames never leave your device, but the model weights are fetched from Google model hosting in the background once the page opens (the small model) or after you tap the download button (the larger one), because they are too large to ship inside the page. The tool is therefore not an offline tool: it needs a connection for that first load, and your browser may cache the files afterwards. Once the model is loaded, every detection runs on your device.
You control what is drawn. A minimum score slider hides weak guesses, and a box limit keeps the most confident ones, up to 100. Detection can run on the full-resolution image or on a copy no larger than 2048 px on its long side, with the boxes scaled back, and the PNG you download is always at full resolution. The page is for trying out object detection and learning how thresholds work. It does not count objects across several images, track objects between video frames or guarantee accuracy. Treat its output as a demonstration, not a measurement.
How to use Online Object Detector
- Choose Image or Webcam, then choose a model. The size of each is shown.
- The small model loads by itself when the page opens, unless Data Saver is on; for the larger one press Download model and confirm. The card says when you can start.
- For an image, drop a picture on the box, tap it to choose one, paste one with Ctrl or Cmd + V, or press Take a photo on a phone.
- For the webcam, press Start camera and allow camera access in your browser.
- Read the boxes on the picture and the text list of detections below it, which gives each label, score and box in pixels.
- Move the Minimum score slider (5% to 95%) or change the box limit (1 to 100) to show fewer or more boxes without detecting again.
- For an image, press Download PNG to save the annotated picture at full resolution.
Features
- Detects 80 COCO object classes with a bounding box, a label and a score for each.
- Two models to choose from: Lite (fast) (about 18 MB, faster) and Accurate (about 67 MB, more accurate), with the download size shown first.
- Detects on the full-resolution image or on a copy up to 2048 px, with the boxes scaled back to the original.
- Image input by drop, file choice, paste or the phone camera.
- Webcam mode that draws video and boxes straight to the canvas in an animation loop, so the page is not re-rendered for every frame. Its text summary updates about four times a second.
- A minimum score slider and a limit on the number of boxes, applied instantly to results that are already found.
- A text list of every detection for screen readers, and a description on the canvas itself.
- Download of the annotated image as a PNG at the full size of the original.
- A built-in list of the 80 detectable objects.
- The camera stops when you press Stop camera, switch to Image mode or leave the page.
Worked examples
A street photo with two people, a bicycle, a car and a faint third person gives At a minimum score of 50%: 2 person, 1 bicycle, 1 car. The faint person at 42% is hidden.
Illustrative output. Real scores depend on the picture. Lowering the slider to 40% would show the third person.
The same picture with the box limit set to 2 gives Only the two highest scores remain: 2 person.
The limit keeps the most confident boxes, not the first ones found.
A 4032 by 3024 photo detected on a 2048 px copy, with a dog box at x 256, y 192, 512 by 384 on the copy gives The copy is 2048 by 1536. The box maps back to x 504, y 378, 1008 by 756 on the original.
The PNG is drawn from the original, so it keeps all 4032 by 3024 pixels.
The two models
| Model | Download | Choose it when |
|---|---|---|
| Lite (fast) | about 18 MB | Smallest download and fastest. Misses more small objects. |
| Accurate | about 67 MB | Larger download and slower. Better on small or partly hidden objects. |
The kinds of objects it recognises
| Group | Examples |
|---|---|
| People and animals | person, cat, dog, horse, bird, elephant, teddy bear |
| Vehicles | bicycle, car, motorcycle, bus, truck, train, boat, airplane |
| Street and outdoor | traffic light, stop sign, bench, fire hydrant |
| Furniture and home | chair, couch, bed, dining table, potted plant, clock |
| Food and kitchen | bottle, cup, bowl, banana, pizza, cake, microwave |
| Electronics and everyday | laptop, cell phone, tv, keyboard, backpack, book |
Good to know
About the model download
Your pictures stay on your device, but the model is not built into the page. It is a set of pretrained weights, and they are downloaded from Google model hosting when you press Download model. The size is shown first: about 18 MB for the lite model and about 67 MB for the accurate one. On a metered connection, choose the lite model.
For that reason the tool needs an internet connection at least once, and you should not rely on it offline. The page code that runs the model is part of this site, and the detection itself happens on your device. Your images are not part of the download.
Full resolution or a capped copy
Analysing a 12-megapixel photo at full size costs time and memory, especially on a phone. By default the tool detects on a copy no larger than 2048 px on its long side and scales the boxes back. Choose Full resolution if you want the model to see every pixel. Expect it to be slower, and on a phone it can run out of memory.
Either way the picture on screen and the PNG download are drawn from the original image, at its full size.
COCO-SSD, not YOLO
COCO-SSD is a single-shot detector trained on the COCO dataset. It is small and fast enough to run in a browser, which makes it a good demo, but it is less accurate than larger models, especially for small, partly hidden or unusual objects. It does not detect faces, read text or identify individual people.
Privacy: how it works
- Images and webcam frames are processed in your browser and are never uploaded to Ulkaa or to anyone else. The camera stream is not recorded or stored.
- One download is needed: the pretrained COCO-SSD weights come from Google model hosting after you press Download model (about 18 MB or 67 MB). That request does not include your images. The TensorFlow.js code and the COCO-SSD code are part of this site.
Tips
- Use a clear, well-lit picture where the object takes up a good part of the frame. Small or blurry objects are often missed.
- If you see wrong labels, raise the minimum score. If objects are missing, lower it, but expect more false boxes.
- Try the accurate model if the lite one misses small objects. It takes longer to download and to run.
- In webcam mode, allow camera access when your browser asks. Camera access needs a secure (https) page.
- Stop the camera when you are finished. Detection runs continuously while it is on, which uses processor power and battery.
- On a phone, Take a photo opens the rear camera. To pick from your gallery, tap the drop box instead.
Frequently asked questions
Is it really free?
Yes. There is no price, no trial and no credit card. Ulkaa Tool does the work with your own device, so there is no server bill to pass on to you.
Do I need an account?
No. There is no sign-up and no login. Open the page and use it. An Ulkaa account is not needed for anything on this tool.
Does it run in my browser?
Yes. Your images are processed by code running in your browser tab, on your own device.
Is my data uploaded?
No. Your images are never sent to Ulkaa or to anyone else. You can check this yourself by watching the Network tab in your browser developer tools while you use the tool.
Is anything downloaded?
Yes, once per model: the pretrained COCO-SSD weights, about 18 MB for the lite model or 67 MB for the accurate one, are downloaded from Google model hosting after you press Download model. Your images and camera frames are never uploaded.
Does it work offline?
No, not reliably. The model weights have to be downloaded from Google model hosting, so it needs a connection to load the model. Detection then runs on your device.
Which model should I choose?
Lite (about 18 MB) downloads fast and runs fast, and suits webcam use. Accurate (about 67 MB) is slower and larger but finds more small or partly hidden objects.
What objects can it detect?
The 80 classes of the COCO dataset, including people, vehicles, animals, furniture, food and everyday household objects. The list is on the tool page.
Is it YOLO?
No. It uses COCO-SSD running on TensorFlow.js. A YOLO model is a different architecture and is not used here.
What does the score mean?
It is the model score for that label, from 0 to 100 percent. A higher score means the model is more certain, but it is not a guarantee, and a confident label can still be wrong.
Is my picture shrunk before it is saved?
No. Detection may run on a copy up to 2048 px, but the picture on screen and the PNG download come from the original at full size.
Does it save or upload my webcam video?
No. Frames are processed in the page and drawn on a canvas. Nothing is recorded or sent anywhere.
Why did it miss an object in my photo?
The object may not be one of the 80 classes, may be small or partly hidden, or may score below your minimum score. Try lowering the slider, using the accurate model or a clearer photo.