COMP9517 Group Project - Insect Detection and Classification

AgroPest-12 dataset insect detection and classification using multiple computer vision methods.

Project Structure

LaserFocus999/
├── data/AgroPest-12/       # Dataset (not in git)
│   ├── train/              # Training set (11,502 images)
│   │   ├── images/
│   │   └── labels/         # YOLO format (.txt)
│   ├── valid/              # Validation set (1,095 images)
│   └── test/               # Test set (546 images) - ONLY for final evaluation
│
├── classical_sift/         # Classical ML: SIFT + BoW + SVM
│   ├── src/                # Feature extraction and classification
│   ├── configs/            # Configuration files
│   ├── classical-results/  # Evaluation results
│   └── README.md           # Detailed documentation
│
├── faster-rcnn/            # Faster R-CNN (Detectron2)
│   ├── configs/            # Model configurations (baseline + improvements)
│   ├── scripts/            # Training, evaluation, visualization scripts
│   ├── src/                # Dataset registration, custom models
│   ├── faster-rcnn-results/ # Test set evaluation results
│   ├── TRAINING_RECORD.md  # Detailed experimental logs
│   └── README.md           # Complete documentation
│
├── yolov8s/                # YOLOv8s implementation
│   ├── ass1.ipynb          # Training and evaluation notebook
│   ├── data.yaml           # Dataset configuration
│   ├── yolov8s_val_summary.csv        # Validation metrics
│   └── yolov8s_val_per_class_AP.csv   # Per-class AP results
│
├── yolo11s/                # YOLOv11s implementation (latest YOLO)
│   ├── 9517_project_yolo11s.ipynb     # Training and evaluation notebook
│   ├── data.yaml           # Dataset configuration
│   └── results.csv         # Evaluation results
│
├── yolo/                   # Original YOLO experiments
│
├── comparison/             # Cross-method evaluation and comparison
│   ├── scripts/            # Unified evaluation scripts
│   └── results/            # Comparative results from all methods
│
├── utils/                  # Shared utility functions
├── docs/                   # Documentation and guides
├── figures/                # Figures and visualizations for report
└── detectron2/             # Detectron2 library (dependency)

Setup

1. Clone Repository

git clone <repository-url>
cd LaserFocus999

2. Download Dataset

From Kaggle: https://www.kaggle.com/datasets/rupankarmajumdar/crop-pests-dataset

Using Kaggle CLI:

pip install kaggle
# Setup kaggle.json in ~/.kaggle/
kaggle datasets download -d rupankarmajumdar/crop-pests-dataset --unzip

Place extracted files in data/AgroPest-12/

3. Create Your Branch

git checkout -b YourName/MethodName

Important Notes

Dataset

• Format: YOLO format (images + labels)
• Splits: Pre-defined train/valid/test - DO NOT modify or create your own
• YOLO team: Use directly without conversion
• Other methods: Convert to required format (e.g., COCO for Detectron2)

Git Workflow

# Work on your branch
git add .
git commit -m "Your commit message"
git push origin YourName/MethodName

# Create Pull Request to merge into main

What NOT to Commit

Already configured in .gitignore:

• Dataset files (data/, images)
• Trained models (*.pth, *.h5, models/)
• Results and outputs (results/, output/)
• API keys (.env, .claude/)

Required Evaluation Metrics

All methods must report on the test set:

Detection:

• Mean Average Precision (mAP)
• mAP@0.5, mAP@0.75

Classification:

• Precision, Recall, F1 Score (per-class and average)
• Accuracy
• AUC

Efficiency:

• Training time
• Inference time (FPS)
• GPU memory usage

Implemented Methods

This project compares four different computer vision approaches for insect detection and classification:

1. Classical ML: SIFT + BoW + SVM

Location: classical_sift/

Approach:

• Feature extraction using SIFT (Scale-Invariant Feature Transform)
• Bag-of-Words (BoW) feature encoding with K-means clustering
• SVM (Support Vector Machine) classifier for classification
• Selective Search for region proposal (detection)

Key Features:

• No deep learning, handcrafted features
• Interpretable feature representations
• Lower computational requirements

Documentation: See classical_sift/README.md

2. Faster R-CNN (Detectron2)

Location: faster-rcnn/

Approach:

• Two-stage object detector with Region Proposal Network (RPN)
• ResNet-50-FPN backbone for feature extraction
• Pre-trained on COCO dataset, fine-tuned on AgroPest-12
• Multiple experimental configurations tested

Best Model: Cascade RCNN + GIoU loss (43.86% mAP)

Key Features:

• Multi-scale anchors for various insect sizes
• Test-time augmentation (TTA)
• Grad-CAM visualization for model interpretability

Dataset Format: COCO JSON (converted from YOLO)

• Conversion script: faster-rcnn/scripts/yolo_to_coco.py

Documentation: See faster-rcnn/README.md and faster-rcnn/TRAINING_RECORD.md

3. YOLOv8s

Location: yolov8s/

Approach:

• Single-stage real-time object detector
• YOLOv8 small variant optimized for speed-accuracy tradeoff
• End-to-end training on AgroPest-12

Key Features:

• Native YOLO format support (no conversion needed)
• Fast inference speed
• Jupyter notebook-based workflow

Files:

• Training notebook: ass1.ipynb
• Results: yolov8s_val_summary.csv, yolov8s_val_per_class_AP.csv

4. YOLOv11s

Location: yolo11s/

Approach:

• Latest YOLO architecture (November 2024 release)
• Improved feature extraction and detection head
• State-of-the-art single-stage detector

Key Features:

• Enhanced performance over YOLOv8
• Native YOLO format support
• Modern architecture improvements

Files:

• Training notebook: 9517_project_yolo11s.ipynb
• Results: results.csv

Method-Specific Notes

Classical SIFT-BoW-SVM

• Detection uses Selective Search instead of bounding-box supervised training
• Feature vocabulary size: configurable K-means clusters
• See classical_sift/README.md for configuration details

YOLO Methods (YOLOv8s, YOLOv11s)

• Dataset already in YOLO format - use directly
• No conversion needed
• Training via Jupyter notebooks
• Results exported to CSV format

Faster R-CNN / Detectron2

• Requires YOLO → COCO format conversion
• Conversion script: faster-rcnn/scripts/yolo_to_coco.py
• Supports command-line training and evaluation
• Multiple experimental configurations available

Deliverables

1. Video: Max 10 min, MP4, <100MB
2. Report: Max 10 pages, IEEE format, PDF, <10MB
3. Code: ZIP, <25MB (no models/data)

Method Comparison Summary

Method	Type	Architecture	Dataset Format	Training Time	Inference Speed	Best mAP	Key Advantage
SIFT-BoW-SVM	Classical ML	Handcrafted features + SVM	YOLO (direct)	Fast	Fast	TBD	Interpretable, no GPU needed
Faster R-CNN	Two-stage DL	ResNet-50-FPN	COCO (converted)	~50 min	Moderate	43.86%	Best accuracy, Cascade+GIoU
YOLOv8s	Single-stage DL	YOLOv8 small	YOLO (direct)	Moderate	Fast	TBD	Speed-accuracy balance
YOLOv11s	Single-stage DL	YOLOv11 small	YOLO (direct)	Moderate	Fastest	TBD	Latest architecture, real-time

Performance Notes:

• Faster R-CNN achieves highest mAP (43.86%) with Cascade RCNN + GIoU configuration
• YOLO methods prioritize inference speed for real-time applications
• Classical SIFT-BoW-SVM provides baseline and interpretable features
• All methods evaluated on same AgroPest-12 test set (546 images, 12 classes)

Recommendations:

• Best Accuracy: Faster R-CNN (improve3 configuration)
• Best Speed: YOLOv11s
• Best Interpretability: SIFT-BoW-SVM
• Best Balance: YOLOv8s

Resources

• Dataset: https://www.kaggle.com/datasets/rupankarmajumdar/crop-pests-dataset
• Detectron2: https://github.com/facebookresearch/detectron2
• Ultralytics YOLOv8: https://github.com/ultralytics/ultralytics
• YOLOv11: https://docs.ultralytics.com/models/yolo11/