Computer Vision with Deep Learning

Convolution, Padding, Stride and Pooling

The operations inside a CNN, and output sizes

Code: convolution from scratch

out = (n + 2p − k) / s + 1

CNN Architectures

LeNet, AlexNet, VGG, Inception, ResNet, MobileNet and EfficientNet

Code: compare architectures

Transfer Learning and Fine-Tuning

Feature extraction and full fine-tuning

Code: fine-tune ResNet on a new dataset

Data Augmentation

Flips, crops, colour changes, MixUp and CutMix

Code: augmentation pipeline

Object Detection

Two-stage and one-stage detectors, anchors, IoU, NMS and mAP

Code: train a YOLO detector

IoU = |A∩B| / |A∪B|

Semantic and Instance Segmentation

U-Net, Mask R-CNN and the Dice score

Code: segment medical images

Dice = 2|A∩B| / (|A| + |B|)

Promptable Segmentation

Segment Anything and interactive segmentation

Lab: segment with prompts

Vision Transformers

Images as sequences of patches

Code: fine-tune a ViT

Self-Supervised Vision

Contrastive learning and masked image modelling

Code: SimCLR-style training

Video Understanding

3D CNNs, two-stream networks and video transformers

Code: action recognition

3D Vision and Neural Rendering

Depth, stereo, NeRF and Gaussian splatting

Lab: reconstruct a scene

Optical Character Recognition and Document AI

Reading text and layout from documents, including Indian scripts

Code: extract data from a form

Deploying Vision Models

Edge deployment, quantisation and real-time inference

Lab: run a model on a phone or browser

Module Review and Assessment

Coding tasks and a scored quiz

Graded project