Computer Vision Foundations

Week 1

Course Overview & Image Formation

The CV-AI pipeline from camera sensor to prediction, images as 2D discrete signals, and color models from RGB to HSV and LAB.

• The CV-AI Pipeline — from camera sensor to prediction
• Images as 2D discrete signals: pixels, color models & spaces
• App: Autonomous driving & medical imaging — how cameras feed AI

Week 2

Spatial Filtering & Kernels

Convolution as a spatial operation, the sliding kernel intuition, and building an industrial edge-detection pipeline.

• Convolution as a spatial operation — the sliding kernel intuition
• Gaussian blur & edge detection: Sobel, LoG & Canny
• App: Industrial surface inspection & document binarization

Week 3

Feature Detection & Matching

What makes a good feature, Harris corners, SIFT & ORB descriptors, and building a keypoint matching pipeline.

• What is a feature? Harris corners & scale-space theory
• SIFT & ORB: descriptors, invariance & ratio-test matching
• App: AR marker tracking, document scanning & product recognition

Week 4

Image Alignment & Panoramas

Homography estimation with DLT, RANSAC for robust fitting under outliers, and automatic panorama stitching.

• Homography matrix H & Direct Linear Transform (DLT)
• RANSAC: robust estimation under outlier correspondences
• App: Panorama stitching & medical image registration

Week 5

Neural Networks for Vision

From perceptron to MLP, backpropagation and the chain rule, and training a digit classifier from scratch in PyTorch.

• From perceptron to MLP: activation, loss & the chain rule
• Backpropagation & optimization — SGD, Momentum & Adam
• App: Digit & handwriting classification as a vision foundation

Week 6

Convolutional Neural Networks (CNNs)

Conv layers, pooling, receptive fields, and the CNN feature hierarchy — from LeNet to AlexNet across six decades.

• Convolutional layer: filters, stride, padding & feature maps
• Pooling, receptive fields & the CNN feature hierarchy
• App: Image classification from LeNet to AlexNet — a 60-year journey

Week 7

Modern CNNs & Transfer Learning

Skip connections in ResNet, efficient convolutions in MobileNet, and fine-tuning strategies for custom datasets.

• Skip connections (ResNet) & efficient convolutions (MobileNet)
• Transfer learning: pre-trained weights & fine-tuning strategies
• App: Fine-tuning MobileNet for custom product quality control

Week 9

Object Detection Frameworks

Bounding boxes, IoU, anchor design, and a side-by-side comparison of two-stage vs. single-stage detector architectures.

• Bounding boxes, IoU & the leap from classification to localization
• Two-stage (Faster R-CNN) vs. single-stage (YOLO) architectures
• App: Real-time YOLO for traffic monitoring & retail analytics

Week 10

Semantic & Instance Segmentation

Pixel-wise classification with encoder-decoder U-Net, Mask R-CNN for instance masks, and medical image segmentation.

• Pixel-wise classification: the encoder-decoder architecture (U-Net)
• Instance segmentation (Mask R-CNN) & panoptic unification
• App: Tumor segmentation & autonomous driving scene parsing

Week 11

3D Vision & Multi-View Geometry

Camera calibration, stereo disparity-to-depth, epipolar geometry, and Structure from Motion for 3D reconstruction.

• Camera calibration, stereo vision & disparity-to-depth
• Epipolar geometry, fundamental matrix & Structure from Motion
• App: Autonomous vehicle depth perception & drone 3D reconstruction

Week 12

Vision Transformers (ViT)

Scaled dot-product attention, multi-head attention, patch embeddings, and Swin's hierarchical shifted-window approach.

• Self-attention mechanism: scaled dot-product & multi-head attention
• ViT: patch embeddings; Swin: hierarchical shifted windows
• App: Swin Transformer for satellite & aerial image analysis

Week 13

Generative Models in Vision

VAE latent spaces, adversarial training in GANs, and diffusion model denoising — with medical image synthesis as a case study.

• VAE & GAN: latent space learning & adversarial training dynamics
• Diffusion models: forward noise process & learned reverse denoising
• App: Medical image synthesis for data augmentation

Week 14

Video Tracking & Edge AI

Optical flow with Lucas-Kanade, DeepSORT multi-object tracking, model compression, and ONNX/TensorRT edge deployment.

• Optical flow (Lucas-Kanade) & multi-object tracking (DeepSORT)
• Model optimization: pruning, quantization & knowledge distillation
• App: Edge deployment with TensorRT & ONNX for embedded vision