Extracting structure from images.
11 topics · 23 curated works
No prior grounding assumed.
Lecture 1 | Introduction to Convolutional Neural Networks for Visual Recognition
Stanford University School of Engineering · 2017
Opens Stanford's CS231n course by tracing computer vision from early edge-detection theory to the ImageNet moment, framing the rest of the course as…
Assumes you know the vocabulary.
Computer Vision: Algorithms and Applications
Richard Szeliski · 2010
Walks the working vision pipeline — features, alignment, stereo, motion, recognition — keeping each piece of mathematics next to the application that…
+2 more at this level
Primary sources and full treatments.
ImageNet Classification with Deep Convolutional Neural Networks
Krizhevsky, Sutskever & Hinton · 2012
AlexNet's error rate on ImageNet made deep learning credible and redirected the entire field within a year.
+18 more at this level
12 of 23 works
Stanford University School of Engineering
Opens Stanford's CS231n course by tracing computer vision from early edge-detection theory to the ImageNet moment, framing the rest of the course as the story of why learned features replaced hand-designed ones.
68 minuteslink checked 17 Sept 2026Richard Szeliski
Walks the working vision pipeline — features, alignment, stereo, motion, recognition — keeping each piece of mathematics next to the application that needs it.
link checked 17 Sept 2026Ronneberger, Fischer & Brox
An encoder-decoder with skip connections that segments accurately from very few labelled training images, becoming the default architecture for biomedical segmentation.
8 pageslink checked 17 Sept 2026Redmon et al.
Reframes object detection as a single regression problem solved by one neural network pass, trading a little accuracy for real-time speed.
10 pageslink checked 17 Sept 2026Krizhevsky, Sutskever & Hinton
AlexNet's error rate on ImageNet made deep learning credible and redirected the entire field within a year.
9 pageslink checked 17 Sept 2026Girshick, Donahue, Darrell & Malik
Shows that region proposals combined with CNN features (R-CNN) beat hand-engineered detectors by a wide margin, launching the region-based detection lineage.
21 pageslink checked 17 Sept 2026Simonyan & Zisserman
Splits video action recognition into a spatial CNN stream on individual frames and a temporal CNN stream on stacked optical flow, showing motion has to be modelled explicitly for a CNN to recognise actions well.
9 pageslink checked 17 Sept 2026He, Zhang, Ren & Sun
Depth had started making networks worse rather than overfit; learning residuals instead of full mappings makes hundreds of layers trainable and fixes the degradation.
link checked 17 Sept 2026Long, Shelhamer & Darrell
Replaces the fully-connected classification layers of a CNN with convolutions and upsampling, making dense pixel-level prediction trainable end to end.
10 pageslink checked 17 Sept 2026Fischer et al.
Argues optical flow can be learned end to end by a CNN trained on synthetic data, rather than solved by the hand-built energy minimisation methods that preceded it.
9 pageslink checked 17 Sept 2026Henriques, Caseiro, Martins & Batista
Exploits the circulant structure of overlapping image patches to turn correlation-filter tracking into a diagonalised, closed-form problem solvable orders of magnitude faster than earlier trackers.
14 pageslink checked 17 Sept 2026Tran, Bourdev, Fergus, Torresani & Paluri
Argues that 3x3x3 convolutions applied across stacked video frames (C3D) learn a single general-purpose spatiotemporal feature that transfers across action recognition, scene, and object tasks.
9 pageslink checked 17 Sept 2026