Appendix A β History of Computer Vision
βTo understand where computer vision is going, you must understand where it came from β and the surprising number of times the field has reinvented itself.β
Pre-Deep-Learning Era (1959β2011)
The Neurobiological Origins (1959β1970s)
Planned content: Hubel & Wieselβs Nobel Prize work (1981) on simple and complex cells in the cat visual cortex β the biological basis for edge detectors and hierarchical feature processing. Block world (Larry Roberts, 1963): first CV thesis. DARPA autonomous vehicle (1970s). The βSummer Projectβ (1966) that thought CV would be solved by the end of the summer.
Feature Descriptor Era (1980β2011)
Planned content: SIFT (Lowe, 2004): scale-invariant feature transform, robust keypoint matching. SURF: faster SIFT. HOG (Dalal & Triggs, 2005): histograms of oriented gradients for pedestrian detection. Viola-Jones face detector (2001): Haar features + AdaBoost + integral images β first real-time face detection. Deformable Part Models (DPM, Felzenszwalb, 2008). PASCAL VOC challenge beginning (2005).
π Planned diagram: SIFT keypoint detection and description pipeline β scale-space extrema β orientation assignment β 128D descriptor.
Shallow Machine Learning for CV (1990sβ2011)
Planned content: SVM on HOG/SIFT features. Bag of Visual Words (BoVW): VQ of SIFT β histogram β SVM. Fisher Vector: improved BoVW encoding. Principal Component Analysis (PCA) for face recognition (Eigenfaces, Turk & Pentland, 1991). The kernel trick for non-linear CV.
The Deep Learning Revolution (2012β2015)
AlexNet and ImageNet (2012)
Planned content: Krizhevsky, Sutskever, Hinton. 8-layer CNN, ReLU, dropout, GPU training. 15.3% top-5 error vs. 26.2% for runner-up. Changed everything. The role of ImageNet challenge in catalyzing progress. GPU training being the key enabler.
The Feature Learning Explosion (2013β2015)
Planned content: ZFNet (2013): visualizing CNN features. VGGNet (2014): 3Γ3 convolutions everywhere. GoogLeNet/Inception (2014): wider network, 1Γ1 convolutions, auxiliary classifiers. Batch Normalization (2015). R-CNN (2013) β Fast R-CNN (2015) β Faster R-CNN (2015): detection revolution.
The Architecture Era (2015β2019)
π Planned diagram (timeline): Architecture evolution β ResNet (2015) β DenseNet (2016) β MobileNet (2017) β NASNet (2018) β EfficientNet (2019).
ResNet (2015)
Planned content: He et al. Skip connections solving vanishing gradients. Won ILSVRC with 3.57% error. 152 layers β deeper than any prior network. Highway networks as predecessor.
Semantic Segmentation Breakthrough (2015β2016)
Planned content: FCN (2014 arxiv, 2015 CVPR): first end-to-end segmentation. U-Net (2015): medical imaging watershed. DeepLab v2 (2016): ASPP.
Object Detection Explosion (2015β2019)
Planned content: YOLOv1 (2015): detection as regression, 45 FPS. SSD (2016): multi-scale. FPN (2017): feature pyramids. Mask R-CNN (2017): instance segmentation. RetinaNet (2017): focal loss. YOLOv3 (2018). CornerNet (2018). FCOS (2019).
The Transformer Era (2020β2023)
Vision Transformer (2020)
Planned content: Dosovitskiy et al. ViT: applying BERT-style transformers to image patches. Requires large data (JFT-300M). DeiT: ImageNet-only training. Swin (2021): hierarchical windows, SOTA on all dense prediction tasks.
Self-Supervised Learning Revolution (2020β2022)
Planned content: MoCo (2020), SimCLR (2020), BYOL (2020), MAE (2021), DINO (2021). Labels no longer required for good features. Transformers + SSL = new paradigm.
CLIP and Multimodal Era (2021)
Planned content: CLIP (OpenAI, 2021): contrastive language-image pre-training. Zero-shot classification. Foundation for all modern VLMs. DALL-E, GLIDE, DALL-E 2, Stable Diffusion.
The Foundation Model Era (2023βPresent)
Planned content: SAM (2023): segment anything. DINOv2 (2023): universal visual features. GPT-4V (2023): multimodal LLM. Gemini 1.5 (2024): native multimodal. SAM 2 (2024): video segmentation. Florence-2 (2024): unified vision. The convergence of vision and language.
π Planned diagram (full timeline): Comprehensive CV history timeline β key papers, paradigm shifts, and benchmark milestones from 1959 to 2026.
Last updated: May 2026