the models only see what their dataset taught them. upload anything, and find what they can’t.
⊞
drop an image or click to choose
a photo, a rendering, a collage, anything
run a detector
pick a model and let it find what it can
base model
Five YOLO variants (11n/s/m, 12n/s) run in your browser via ONNX Runtime (WASM); COCO-SSD runs via TensorFlow.js. The bigger models (11m especially) see more but download slower. Everything runs locally, so your image never leaves your device. It’s only added to the shared dataset when you save it. All were trained on COCO (80 everyday classes), so none has a word for “building.” that blind spot is the point.
what the machine sees
the detector only recognizes what its dataset taught it. everything it has no word for dissolves into noise. slide the confidence to watch the edge of the recognizable move.
model
confidence.35
review the labels
it zooms into each box one at a time. keep, relabel, or reject, and it walks to the next
review each box, then continue
add to the archive
these labels become part of our shared archive. enough of them, and we can retrain the models to see what they currently can’t.
saving these labels:
✓ saved
the archive we’re building
every image people have labeled, gathered into one archive. the data is the design, and eventually, enough of it could fine-tune the models to see what they currently can’t.
this dataset is empty
nothing has been added yet. go to “detect something,” label an image, and add it here.