Semester 3 · Computer vision
Deep Visual Recognition
https://epochschool.com/subjects/deep-visual-recognition
Learn how modern vision systems see: the CNN and Vision Transformer architectures behind them, the training recipes and transfer strategies that make them work on real data, and the detectors and segmenters that find and outline every object in an image.
You extend recognition to video and tracking, learn representations from unlabelled images with self-supervised learning, check what a model is really looking at, and compress it for a CPU-bound device, measuring size, latency and accuracy honestly. Every method is derived, costed and built from scratch in PyTorch on 8 x 8 digits and synthetic 32 x 32 scenes, so all five labs run on a single CPU core with no downloaded or pretrained models.
- Prerequisites
- Required: Finish this subject before you start. The subject builds directly on it.Recommended: Helpful background. You can start without it if you are willing to fill the gaps.
What you’ll be able to do
- LO1Analyse and compare CNN architectures (AlexNet, Inception, ResNeXt, DenseNet, MobileNet, EfficientNet, ConvNeXt) by computing their parameters, MACs and arithmetic intensity, and choose one for a given accuracy, latency and memory budget.
- LO2Explain, implement and analyse Vision Transformers (patch embedding, 2-D positions and their interpolation, windowed and hierarchical attention, distillation tokens) and compare their inductive biases, cost and data needs with those of CNNs.
- LO3Design and justify vision training recipes (augmentation including Mixup and CutMix, stochastic depth, weight averaging, imbalance handling) and transfer strategies (linear probe, fine-tuning with layer-wise rates, batch-norm adaptation, few-shot prototypes), and test them with controlled, multi-seed experiments.
- LO4Implement and evaluate object detectors: IoU and GIoU, anchors and target assignment, box encoding, the focal loss, non-maximum suppression and AP/mAP, and compare two-stage, one-stage, anchor-free and set-based designs.
- LO5Implement and evaluate semantic, instance and panoptic segmentation: transposed convolution and upsampling, encoder-decoder networks with skips, cross-entropy and Dice losses, mIoU and panoptic quality.
- LO6Implement and evaluate self-supervised representation learning (contrastive InfoNCE/NT-Xent, non-contrastive SimSiam/BYOL, masked autoencoding) with linear and k-NN probes against honest baselines, and diagnose representation collapse.
- LO7Build and analyse video recognition and tracking models: temporal pooling, 3-D and (2+1)D convolutions, two-stream and space-time attention designs, and tracking-by-detection with Hungarian assignment, including their cost and evaluation.
- LO8Compute and critically evaluate visual explanations (gradient saliency, SmoothGrad, integrated gradients, CAM and Grad-CAM) with completeness checks, sanity checks and deletion curves, and use them with interventions to detect shortcut learning.
- LO9Compress vision models for deployment with batch-norm folding, affine int8 quantisation with calibration and integer arithmetic, quantisation-aware training, pruning and knowledge distillation, measuring accuracy, size and CPU latency honestly.
Chapters
The chapter and lab list for this subject is being finalised and will be published here.
Before you start
Required
Finish this subject before you start. The subject builds directly on it.
Sequence Models & Transformers
Recommended
Helpful background. You can start without it if you are willing to fill the gaps.
Optimisation & Training at Scale
Image Formation and Classical Computer Vision
Where it leads
Subject 2 of 4 in the Computer vision track.
Previous: Image Formation and Classical Computer VisionNext: Deep Generative Models
Deep Generative Models
Vision-Language, 3D Vision and World Models
Capstone Project
How it’s taught
You read short sections written for the page. Each ends with a check, and in-browser exercises let you practise. Graded labs run on your own cloud lab machine, CPU-friendly.
No videos. No live classes. No AI tutor.