Programming Assignments for Computer Vision (Hindi)

परिचय

यह लेख Computer Vision के व्यावहारिक प्रोजेक्ट और Programming Assignments के लिए एक संरचित मार्गदर्शिका है। B.Tech Data Science / Computer Vision लैब्स के लिए बनाए गए असाइनमेंट step-by-step, learning outcomes, datasets, evaluation rubrics और implementation hints (Python/OpenCV, PyTorch/TensorFlow) के साथ दिए गए हैं। हर असाइनमेंट का उद्देश्य सिद्धान्त को व्यावहारिक रूप में समझना और industry-ready code-writing skill विकसित करना है।

इस पेज का उद्देश्य (Learning Goals)

• Core CV algorithms को implement और experiment करना (filters, transforms, feature detectors, tracking)।
• Modern deep learning pipelines (CNN, object detection, segmentation, autoencoders) से practical अनुभव प्राप्त करना।
• End-to-end project design: dataset → preprocessing → model → evaluation → report
• Code quality, reproducibility और performance optimization सिखना

Suggested Assignments (Scaffolded from Easy → Hard)

Assignment 1 — Image Transforms & Filters (Beginner)

• Tasks: Implement convolution, Gaussian blur, median filter, Sobel edge detector, histogram equalization.
• Dataset: A few sample images (grayscale + color)
• Deliverables: Python scripts using OpenCV, before/after images, short report describing parameters and effect.
• Learning Outcome: Understand spatial filtering and contrast enhancement.

Assignment 2 — Color Processing & Segmentation (Easy)

• Tasks: Convert RGB→HSV/YCbCr, implement colour thresholding, K-means colour clustering, simple watershed segmentation.
• Dataset: Fruits/traffic-sign images
• Deliverables: Segmented outputs, accuracy against simple ground truth masks.

Assignment 3 — Feature Detection & Matching (Intermediate)

• Tasks: Detect keypoints (Harris/SIFT/ORB), compute descriptors, match between image pairs, RANSAC for homography estimation, image stitching.
• Dataset: Overlapping photo sequences (own captures or standard datasets)
• Deliverables: Matched visualizations, stitched panorama, explanation of failure-cases.

Assignment 4 — Optical Flow & Tracking (Intermediate)

• Tasks: Implement Lucas-Kanade sparse optical flow, Farneback dense flow; build a simple object tracker (KLT + Kalman).
• Dataset: Short video clips (walking person, moving objects)
• Deliverables: Tracking visualization, performance (FPS, identity switches) analysis.

Assignment 5 — Image Classification using CNN (Advanced)

• Tasks: Train a CNN (e.g., small ResNet) on CIFAR-10 or custom dataset; use data augmentation, evaluate with confusion matrix, implement transfer learning.
• Deliverables: Training scripts (PyTorch/TensorFlow), model checkpoints, accuracy plots, written analysis.

Assignment 6 — Object Detection (Advanced)

• Tasks: Use a pre-trained YOLO/SSD model; fine-tune on a small custom dataset; implement NMS and evaluate mAP.
• Deliverables: Detector with inference script, dataset annotation format, mAP/P-R curves.

Assignment 7 — Segmentation & Autoencoders (Advanced)

• Tasks: Implement U-Net for semantic segmentation or a convolutional autoencoder for denoising and anomaly detection.
• Deliverables: Model, qualitative results (masks), IoU/Dice metrics, ablation study on augmentation.

Assignment 8 — Final Project (Capstone)

• Options: Real-time hand-gesture control, multi-object tracker for CCTV, defect detection on manufacturing line, sign-language translator (skeleton + transformer)
• Deliverables: Full codebase, README for reproducibility, video demo, report with experiments and deployment notes.

Implementation Hints & Boilerplate

• Prefer Python3, OpenCV for classical tasks.
• Use PyTorch/TensorFlow for deep learning. Provide requirements.txt and seed control for reproducibility.
• Structure project: data/, src/, notebooks/, checkpoints/, results/
• Use GPU for training; provide options for CPU fallback.
• Include unit tests for small utility functions (e.g., image IO, normalization).

Evaluation Rubric (Sample)

Datasets & Annotation Tools

• Use public datasets: CIFAR, MNIST, COCO (for detection), PASCAL VOC, DAVIS (segmentation), KITTI (tracking)
• Annotation tools: LabelImg (bbox), CVAT (segmentation & video), VIA

Academic Integrity & Collaboration

• Students should submit individual code but may collaborate in design—acknowledge contributions.
• Plagiarism detection via code similarity tools recommended.

Optional Extensions & Extra Credit Ideas

• Quantize / prune models and report latency improvements
• Deploy model as REST API (Flask/FastAPI) and create a small web demo
• Implement real-time demo on Raspberry Pi / Jetson Nano

Conclusion

ये programming assignments theoretical knowledge को practical expertise में बदलने के लिए डिज़ाइन किए गए हैं। प्रत्येक असाइनमेंट में clarity of objectives, measurable deliverables और reproducible codebase होना चाहिए — ताकि आप industry-ready computer vision engineers बन सकें।