COL828-Advanced-Computer-Vision

Course assignments of COL828:- Advanced Computer Vision course at IIT Delhi under Professor Chetan Arora

This repository contains assignments that explore state-of-the-art vision-language models and prompt learning techniques for medical image classification and detection tasks.

---

📁 Assignments

Assignment 1: Implant Classification

• Applied vision-language models (CLIP, ViT) for medical implant classification on Orthonet and Pacemakers datasets.
• Implemented and compared multiple approaches:
  • Zero-shot CLIP inference without training
  • Fine-tuning with ImageNet21K, CLIP, and DINOv2 backbones
  • Prompt learning methods: CoOp (Context Optimization), CoCoOp (Conditional CoOp), and MaPLe (Multi-modal Prompt Learning)
• Evaluated models using accuracy, F1-score, precision, and recall metrics.
• Demonstrated the effectiveness of prompt learning over traditional fine-tuning for few-shot medical image classification.

---

Assignment 2: Cancer Detection in Mammography

• Performed object detection on mammography images to identify benign and malignant tumors using Grounding DINO.
• Explored zero-shot detection capabilities of pre-trained Grounding DINO without additional training.
• Implemented prompt learning techniques (CoOp, CoCoOp) to improve detection by learning optimal text prompts.
• Applied FixMatch semi-supervised learning to leverage unlabeled data for enhanced model performance.
• Evaluated detection quality using mAP (mean Average Precision) and conducted qualitative analysis with bounding box visualisations.

---

Each assignment demonstrates the application of cutting-edge vision-language models and prompt engineering techniques to challenging medical imaging tasks, combining transfer learning, few-shot learning, and modern detection frameworks.