New 91.6% accuracy on TrashNet

See what the AI sees

Drop a waste photo, watch it get sorted and see what the model looked at

ResNet50

Drop a waste image or click to upload

PNG or JPG, up to 10 MB, 224×224 recommended

JPG / PNG

Original

Grad-CAM

lowhigh

Prediction Plastic

94%confidence

Cardboard

Glass

Metal

Paper

Plastic

Trash

Model	Acc	F1	AUC
ResNet50 live	91.6%	0.916	0.99
Ensemble	91.0%	0.911	0.99
MobileNetV2	84.2%	0.847	0.97
EfficientNet-B0	80.5%	0.806	0.97

The hardest class is trash, with only 137 images in the dataset. A weighted sampler and label smoothing lift its recall well above what a naive model reaches.