About Image Classifier
This image classifier is free, needs no sign-up and runs in your browser, so your pictures are never uploaded. Add a picture, and it lists the most likely labels with a score for each, for example golden retriever, tennis ball or espresso. Slide the labels control to show from one to ten.
It uses the MediaPipe Image Classifier with EfficientNet-Lite0, an efficient network trained on the ImageNet dataset. The model file states an Apache-2.0 licence in its metadata. The model file is about 5.4 MB (the int8 version) and the MediaPipe runtime adds about 13 MB, so the card shows about 19 MB in total. The runtime is served by tool.ulkaa.com itself and only the model file comes from Google model storage, once, and your browser may cache both. Your pictures never leave your device.
The model knows 1,000 everyday classes: many animal breeds, vehicles, foods, tools, clothes, buildings and household objects. It always answers from that list, which has no class for most people, brands, logos, text or abstract scenes. A picture of something outside the list gets the nearest match with a low score, so read the score as well as the label.
The scores come from the model confidence and add up to at most 100 percent across all 1,000 classes. A top score of 90 percent means the model is very sure among its classes, not that the label is certain to be right.
How to use Image Classifier
- The small model starts loading in the background when the page opens, unless Data Saver is on, and the card shows its progress and tells you when you can start.
- Add a picture by dropping it, tapping the box, pasting it or using the phone camera.
- Read the labels and scores in the result card. The bar next to each shows its score.
- Move Labels to show to see more or fewer guesses without running the model again.
- Press the copy button to copy the labels and scores as text.
- If the top score is low, try a closer picture of one object on a plain background.
Features
- Top labels with percentage scores and bars, from one to ten.
- 1,000 ImageNet classes, including animals, vehicles, food, tools and household objects.
- EfficientNet-Lite0 int8: about 5.4 MB, quick on a phone.
- Image input by drop, file choice, paste or phone camera.
- Pictures are shrunk to at most 1024 px before classifying, which does not change the result for a 224 px model.
- Copy the labels as plain text.
- The label list changes without running the model again when you change the count.
- Nothing is uploaded: the model and the runtime are the only downloads.
Worked examples
A photo of a labrador on a lawn gives Labrador retriever 81.2%, golden retriever 9.6%, ...
Illustrative of the format. Run your own picture for real values.
A photo of a person on a bicycle gives Mountain bike or bicycle-built-for-two, with a modest score
There is no class for a person, so the label describes the bike, not the rider.
A screenshot of a web page gives A low top score, such as web site, menu or comic book
A low top score is the model telling you the picture is not a good fit for its classes.
Good to know
Classification is not detection
A classifier gives labels for the whole picture. It does not say where an object is or how many there are. For boxes around several objects use the object detector, and for faces use the face detector.
Why labels can look odd
ImageNet has some very fine classes, such as dozens of dog breeds, and very few for people or scenes. A picture of a beach may be labelled seashore or sandbar, and a picture of a laptop on a desk may be labelled desk or notebook. That reflects the training data, not a fault in the picture.
Privacy: how it works
- Pictures are processed in your browser and are never uploaded to Ulkaa or to anyone else.
- Two downloads are needed: the EfficientNet-Lite0 model (about 5.4 MB) from Google model storage, and the MediaPipe runtime (about 13 MB) from tool.ulkaa.com itself. Neither request contains your pictures.
Tips
- Put one clear object in the middle of the picture for the best result.
- Check the score. Under 30 percent means a weak match.
- Crop to the object before you add the picture if it is small in the frame.
- Use the labels to write alt text, then edit them into a sentence.
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 pictures are processed by code running in your browser tab, on your own device.
Is my data uploaded?
No. Your pictures 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.
What can it recognise?
The 1,000 ImageNet classes: many animals and dog breeds, vehicles, foods, tools, clothes and household objects. It has no class for most people, logos or text.
Which model does it use?
MediaPipe Image Classifier with EfficientNet-Lite0 (int8), from Google, with an Apache-2.0 licence stated in the model file.
What does the score mean?
It is the model confidence for that class among its 1,000. A high score means a strong match, but it does not prove the label is right.
Why is my picture labelled something strange?
The model must pick from its list. If your subject is not on it, you get the nearest match with a low score.
Can it tell me how many objects there are?
No. It labels the whole picture. Use the object detector to find and count objects.
What is downloaded, and from where?
The 5.4 MB model comes from Google model storage, and the 13 MB MediaPipe runtime is served by tool.ulkaa.com itself. Your pictures are never part of those requests.