Image and Video Understanding Project

This project compares two deep learning models for instance segmentation on waste detection: Mask R-CNN (using Detectron2) and YOLOv8.

Project Structure

• MRCNN/: Mask R-CNN implementation using Detectron2

  • Training and evaluation code in main.ipynb
  • Trained models and results in results/

• YOLO/: YOLOv8 segmentation implementation

  • Training and evaluation code in main.ipynb
  • Trained models and results in results/

Dataset

Both models are trained on the TACO (Trash Annotations in Context) dataset with 20 classes of waste objects including:

• Plastic bottles, glass bottles, bottle caps
• Drink cans, paper cups, cartons
• Plastic film, wrappers, straws
• Cigarettes, and other litter items

Models

• Mask R-CNN: ResNet-101 backbone with Feature Pyramid Network
• YOLOv8: Large segmentation model (YOLOv8l-seg)

Results

Training and evaluation results are stored in the respective results/ directories for each model.