point your camera at something. give it a few photos. train a model and watch it recognize the thing you taught it. everything runs in your browser ,you use photos you took yourself.
no leaderboard. no complicated requirements. just build something that works.
no tiers. no points. build it, ship it, get the webcam.
here's a small example of what you're making. turn on your camera, show it rock, paper, and scissors a few times, then hold one up and see what it predicts.
eight steps. about 3 hours total.
pick something you can easily photograph - keep it to 2-4 classes.
pick something you can photograph easily. some ideas:
keep it to 2-4 classes. more classes means more photos and more training.
~30-50 photos per class, varied angle & lighting.
take around 30 to 50 photos per class using your laptop webcam or phone. change the angle and lighting a little - don't take all 50 photos from exactly the same position. this will probably take around 30 to 40 minutes.
one folder per class - or upload straight into teachable machine.
put each class into its own folder, for example biscuit/ and luna/. if you're using teachable machine, you can upload the photos directly into each class instead.
upload to teachable machine, hit train, done in minutes.
open teachable machine. create an image project, add your classes, upload your photos, and train it. training happens in your browser - you don't need a GPU or anything installed.
check the live preview, retrain if it's getting confused.
teachable machine gives you a live webcam preview with the prediction for each class. try each thing you trained it on. if it's getting confused, take some more photos and train it again.
export as tensorflow.js - model.json plus weight files.
once you're happy with it, export the model as tensorflow.js. you should get a model.json file and some weight files. these can run directly in a browser.
~15-20 lines of JS connecting the prediction to a reaction - software-only is fine.
now make the prediction actually do something. we've provided the webcam setup and model loading — your job is to connect the prediction to something. for example:
you don't need to make anything complicated. any of the software-only options above count as a complete answer on their own - hardware is a bonus, not a requirement. this part can be around 15 to 20 lines of javascript.
20-30s demo video, then submit repo + video + teachable machine link.
record a 20 to 30 second video of your project working. show the camera recognizing the thing and the reaction actually happening. then push your project to github and submit your github repo, your demo video, and your teachable machine project.
show us what you made.
some ideas to get you started:
you don't have to build any of these. pick something you actually want to teach your computer.