• Dataset: https://github.com/pedropro/TACO
• Colab Notebook: https://drive.google.com/drive/folders/1K6u6DM8RirkAsTYxmbG-t1f6an99TjuL?usp=sharing
• Streamlit Webapp: https://louisljz-wastewizard2.streamlit.app/
• Approach: Object Detection
• Model: Faster-RCNN

• Get-to-Know annotations.json
• Download Image Data
• Visualize annotations

• Create Dataset Class
• Prepare Transforms
• Split Data into Subsets
• Initiate Data Loaders

• Fine-tune Faster-RCNN Head
• Train Model

• Build Inference Engine

Able to localize various objects in an image, decent classification. However, main weakness is the over-sensitivity of predictions, and latency of inference on CPU.

Over-sensitivity Solution: Group names to material categories