Week 1
Backbones, Self-Supervision and Generative Models
Monday 12 to Friday 16 October. CNNs and Vision Transformers, self-supervised and JEPA pretraining, detection and segmentation, then autoencoders, VAEs, normalizing flows and GANs.
State-of-the-Art Deep Learning Architectures
The backbones the rest of the course builds on: the CNN architectures that won ImageNet, then Vision Transformers, hierarchical variants such as Swin, and how they compare with CNNs.
Foundations of Modern Computer Vision
Learning from unlabelled images: pretext tasks, contrastive learning and self-distillation with SimCLR, BYOL, DINO, DINOv2 and iBOT, then masked autoencoders and JEPA latent prediction.
Object Detection and Segmentation
Finding and outlining objects: IoU, region proposals, the R-CNN and YOLO families, then semantic, instance and panoptic segmentation with FCNs and U-Net.
Flow-Based and Latent Variable Models
Generative models with a likelihood: autoencoders and VAEs, the ELBO and the reparameterization trick, VQ-VAE, then normalizing flows from NICE and RealNVP to Glow and flow matching.
Generative AI for Vision: GANs
Generative adversarial networks: the generator-discriminator game and its objective, training and its failure modes, conditional GANs, and unpaired image-to-image translation with CycleGAN.
Week 2
Diffusion, Foundation Models and Real-World Vision
Monday 19 to Friday 23 October. Diffusion and Stable Diffusion, video and vision foundation models, vision-language models and open-world recognition, then restoration, applications and 3D vision.
Diffusion Models
From noise to images: the forward and reverse processes, the ELBO, training and sampling, latent diffusion, then Stable Diffusion with text conditioning, guidance and customization.
Video Understanding and Foundation Models
Learning from video, from motion to action recognition; then vision foundation models such as CLIP, DINOv2, SAM and Depth Anything, and how to adapt them.
Vision-Language Models
Connecting images and text: captioning, joint embeddings and text-to-image generation, then BLIP, Flamingo, LLaVA, and how multimodal LLMs are trained and evaluated.
Image Classification and Recognition
Recognition beyond a fixed label set: adapting pretrained classifiers, fine-grained recognition, metric learning, retrieval and re-identification, and rejecting images of classes the model has never seen.
Image Restoration and Real-World Applications
Restoring degraded images and vision at work in industry, medicine, Earth observation and transportation, then geometry and 3D: cameras, stereo, depth, structure from motion and SLAM.