BBM418 - Computer Vision Laboratory Assignments (Spring 2024-2025)

This repository contains solutions for the three programming assignments of the BBM418 Computer Vision Laboratory course at Hacettepe University. The projects cover diverse areas in computer vision, including document dewarping, image classification, and object detection.

Assignment 1: Document Dewarping with Edge and Line Fitting

Objective: Correct geometric distortions in scanned documents using classical vision techniques.

Techniques Used:

• Custom implementation of Hough Transform for line detection.
• Robust line fitting with RANSAC.
• Homography-based geometric transformation for perspective correction.
• SSIM (Structural Similarity Index) for evaluation.

Challenges Addressed:

• High-resolution image preprocessing.
• Background noise filtering with morphological operations.
• Adaptive ROI masking and contour-based quadrilateral detection.
• Custom post-processing for consistent document boundary detection.

Result Highlights:

• Successfully dewarped many documents with mild distortions.
• Struggled with complex curvature and background clutter.
• SSIM used for comparison across 6 distortion types.

See detailed analysis in Assignment1/report.pdf.

---

Assignment 2: Food Image Classification with CNNs and Transfer Learning

Objective: Classify images of food using both CNNs built from scratch and pre-trained models.

Part 1: CNNs from Scratch

• Two architectures: Classic CNN and Residual CNN.
• Evaluated multiple learning rates and batch sizes.
• Applied Dropout for regularization.
• Visualization: Accuracy/loss plots and confusion matrices.

Part 2: Transfer Learning

• Used MobileNetV2 for fine-tuning.
• Compared two strategies:
  • Training only the fully connected layer.
  • Training the last two convolutional blocks + FC layer.

Result Highlights:

• Best accuracy from fine-tuning MobileNetV2 (85.82% test accuracy).
• Classic CNN achieved 66.55% accuracy from scratch.
• Transfer learning outperformed all scratch-trained models.

See code and evaluation in Assignment2/ and full discussion in Assignment2/report.pdf.

---

Assignment 3: Object Detection and Counting in Drone Images with YOLOv8

Objective: Detect and count cars in aerial parking lot images using YOLOv8n.

Implementation:

• Trained YOLOv8n under four different freezing strategies:
  • Freeze first 5, 10, 21 blocks, and full training.
• Evaluated with different optimizers and learning rates.
• Used a custom score metric combining precision, recall, and loss.

Metrics:

• Exact Match Accuracy, MSE, Precision, and Recall.
• Best model (yolov8n adam lr=0.001 fulltrain) achieved:
  • 49.00% Exact Match Accuracy
  • 3.61 Mean Squared Error

Visualization:

• Bounding box predictions overlaid on test images.
• Comparative charts and training curves included.

Full results and discussion available in Assignment3/report.pdf.

---

Technologies Used

• Python, NumPy, OpenCV
• PyTorch, torchvision
• Ultralytics YOLOv8
• Jupyter Notebooks
• Kaggle GPU & Google Colab environments