EcoSort is a clean, native Windows application that helps users make informed waste disposal decisions using image-based classification and clear, educational guidance. The app is built around a validated ONNX computer vision model and presents results with confidence-aware messaging to keep the experience transparent and trustworthy.
- Image classification into 12 waste categories using an ONNX ResNet50-based model
- Multiple image input methods (upload, drag-and-drop, and optional camera capture)
- Confidence-aware results with High/Medium/Low bands and recommended next steps
- History and insights (local history of classifications)
- Educational hub with recycling and sustainability guidance independent of the model
- Accessibility-minded UX (keyboard navigation, clear language, support for high contrast)
EcoSort is not a universal garbage classifier. The model always predicts one of a fixed set of 12 categories. For items outside the model’s domain, EcoSort communicates uncertainty and provides educational fallback guidance rather than presenting a misleadingly confident result.
- battery
- biological
- brown-glass
- cardboard
- clothes
- green-glass
- metal
- paper
- plastic
- shoes
- trash
- white-glass
EcoSort translates model confidence into user-friendly bands:
- High:
>= 0.90 - Medium:
>= 0.70and< 0.90 - Low:
< 0.70
For low confidence results, EcoSort presents a gentle disclaimer.
- Input node:
images - Input shape:
[1, 3, 224, 224](NCHW) - Output node:
logits - Output shape:
[1, 12]
- Decode image as RGB
- Resize to 224×224 (bilinear)
- Normalize using ImageNet statistics:
- mean:
[0.485, 0.456, 0.406] - std:
[0.229, 0.224, 0.225] - formula:
(pixel/255 - mean) / std
- mean:
- Tensor layout: NCHW,
float32
- Test accuracy: ~97.09%
- Latency (CPU): ~200–500 ms per image
- Latency (DirectML GPU): ~50–150 ms per image
- Memory footprint (typical): ~200–250 MB resident (runtime + buffers)
EcoSort is designed for integration into a WinUI 3 native Windows app using ONNX Runtime.
- Recommended NuGet:
Microsoft.ML.OnnxRuntime - Recommended execution providers:
- DirectML (GPU) when available
- CPU fallback
- The model is not rotation-invariant
- Performance may degrade on:
- extreme lighting or shadows
- heavily compressed images
- very small or distant items
- Some visually similar classes may be confused (notably shoes vs clothes, and glass color variants)
EcoSort uses ONNX Runtime via Windows AI APIs for on-device inference and is built to provide clear, confidence-aware guidance to support better disposal decisions.