SmartBin

Overview

SmartBin is an AI-powered waste classification system. You upload a photo of any piece of trash and a deep learning model decides which category it belongs to — recycle, compost, landfill, hazardous — fast enough to feel like a tool, not a demo. The backend is a FastAPI service that hot-swaps between eleven pretrained image classification models so you can compare accuracy, latency, and memory footprint side by side.

The models are trained separately in a sibling repo, GarbageClassfication, which owns the dataset pipeline, training scripts, and evaluation tools. SmartBin itself is the inference + serving layer.

Supported models

A single dropdown swaps between these, each loaded on demand:

• DenseNet (densenet201)
• EfficientNet (efficientnet_b7)
• GoogLeNet
• LeNet
• MobileNetV2 / MobileNetV3 (mobilenet_v2 / mobilenet_v3_large)
• ResNet (resnet50)
• AlexNet
• ShuffleNetV2 (shufflenet_v2_x2_0)
• VGG
• RegNet (regnet_x_32gf)

Each model runs on whatever hardware is available — CUDA first, then MPS on Apple Silicon, CPU as a fallback — with no configuration.

Stack

• Backend — FastAPI with an /api/predict endpoint that accepts an image upload and a model name, returns the top-k predictions with softmax probabilities.
• Model loading — PyTorch modules loaded lazily from a weights/ directory. Weights are hosted on Google Drive (too large for git) and fetched once at deploy time.
• Frontend — Next.js + TypeScript, drag-and-drop upload with live prediction feedback and a model picker for comparison runs.
• Deployment — Dockerized for portability, AWS-hosted at smrtbin.com.

Notable bits

• Eleven models, one endpoint — the picker turns SmartBin into an impromptu benchmarking tool. Point the same bottle at each model and watch which ones disagree.
• Hot-swappable weights — models load lazily, so the container boots fast and only pays the memory cost for whichever model is actually in use.
• Hardware-adaptive — same code path on CPU, CUDA, and MPS thanks to PyTorch’s torch.device auto-detection.