GEO5017 Project Group 9 Waste Detection in the Built Environment

Overview This project implements a waste detection pipeline for street-view images using transfer learning. Our final reported workflow used a single-model ResNet-50 setup with a labeled folder-based dataset split into train, val, and test. The main pipeline is in waste_detection_resnet50_v2.py.

Dataset structure dataset/ ├── train/ <- (2016-2019) │ ├── waste/ │ └── no_waste/ ├── val/ <- (2020-2021) │ ├── waste/ │ └── no_waste/ └── test/ <- (2022-2023) ├── waste/ └── no_waste/

Environment

• Python 3.11
• PyTorch
• torchvision
• numpy
• matplotlib
• pillow
• scikit-learn

Hardware used

• NVIDIA RTX 1000 Ada Generation GPU
• CUDA 12.8

Install packages pip install torch torchvision numpy matplotlib pillow scikit-learn pip install ultralytics opencv-python

Main settings

• Backbone: ResNet-50
• Batch size: 256
• WeightedRandomSampler enabled
• Two-phase training:
  1. Train classification head
  2. Fine-tune last block

The script saves checkpoints, evaluates on the test set, and can generate Top-100 detections. :contentReference[oaicite:1]{index=1}

Run python waste_detection_resnet50_v2.py

Main outputs

• checkpoints/
• learning_curves.png
• confusion_matrix.png
• roc_curve.png
• top100/
• top100_annotated/

Final workflow used Our final submission used a single-model test-only workflow. Images in dataset/test were ranked by predicted waste confidence, and the Top 100 highest-confidence images were selected as the final result set. This was logically done only on the test set.

Submission contents The final submission archive should contain:

• Report
• Source code (to be reached through link GitHub in report)
• README / ReadMe.txt
• Top 100 waste detections as images
• A CSV file containing some extra information

According to the project brief, the required results file is the Top 100 detections as images. :contentReference[oaicite:2]{index=2}

Main file

• waste_detection_resnet50_v2.py

Optional helper files

• namestrip.py
• singlemodelscore.py
• ensemble.py

Group Group: [09] Members:

• Henryk
• Adriano
• Mathijs