Computer Vision

Computer Vision (CV) is a field of AI that enables machines to interpret and understand visual information from images and videos. It bridges the gap between raw pixel data and high-level understanding.

Overview

graph TD
    A[Computer Vision] --> B[Image Classification]
    A --> C[Object Detection]
    A --> D[Segmentation]
    A --> E[Image Generation]
    A --> F[Vision-Language]
    
    B --> B1[CNNs]
    B --> B2[ViT]
    B --> B3[Transfer Learning]
    
    C --> C1[YOLO]
    C --> C2[R-CNN Family]
    C --> C3[SSD]
    
    D --> D1[Semantic]
    D --> D2[Instance]
    D --> D3[Panoptic]
    
    E --> E1[Diffusion Models]
    E --> E2[GANs]
    
    F --> F1[CLIP]
    F --> F2[SAM]

Key Paradigms

Traditional CV Pipeline

Image → Preprocessing → Feature Extraction (SIFT, HOG) → Classifier (SVM) → Output

Deep Learning CV Pipeline

Image → Neural Network (end-to-end) → Output

Modern CV Pipeline (Foundation Models)

Image + Text Prompt → Foundation Model (CLIP, SAM, DINO) → Output

Evolution of CV

Core Tasks

1. Image Classification

Assigning a label to an entire image. See classification.md.

2. Object Detection

Locating and classifying objects with bounding boxes. See object-detection.md.

3. Segmentation

Pixel-level understanding of images. See segmentation.md.

4. Image Generation

Creating new images from distributions. See diffusion.md.

5. Vision-Language

Connecting visual and textual understanding. See clip.md and sam.md.

Essential Concepts

Convolution Operation

The fundamental building block of CNNs:

Input (H×W×C) ⊛ Kernel (K×K×C) → Output (H'×W'×C')

Output[i,j] = Σ Σ Input[i+m, j+n] × Kernel[m, n]

Receptive Field

The region of input that affects a particular output neuron. Deeper layers have larger receptive fields.

Feature Maps

Outputs of convolutional filters that capture different visual patterns (edges, textures, shapes).

Interview Questions

1. What is the difference between image classification and object detection? Classification assigns one label to an entire image; detection locates multiple objects with bounding boxes and labels.

2. Why did deep learning outperform traditional CV? Deep learning learns hierarchical features automatically (edges → textures → parts → objects) instead of relying on handcrafted features.

3. What is transfer learning and why is it important in CV? Using pre-trained models (on ImageNet) as starting points. Critical because labeled data is expensive and features transfer well across domains.

4. Explain the bias-variance tradeoff in the context of CNNs. Deeper/wider networks reduce bias but increase variance (overfitting risk). Regularization (dropout, data augmentation) controls variance.

Common Mistakes

• ❌ Not normalizing pixel values (should be [0,1] or standardized)
• ❌ Ignoring data augmentation (leads to overfitting)
• ❌ Using too large models for small datasets (use transfer learning)
• ❌ Not handling class imbalance
• ❌ Confusing validation set with test set

Summary

Computer Vision has evolved from handcrafted features to foundation models. The field now emphasizes pre-training on large datasets and fine-tuning for specific tasks. Understanding CNNs, attention mechanisms, and transfer learning is essential for interviews.

Cross-References

• Neural Networks - Foundation concepts
• Transformers - ViT and attention mechanisms
• GANs - Generative models
• CLIP - Vision-language alignment
• Multimodal Models - Combining vision and language