A Primer to Modern Day Computer Vision
summary
Computer vision history moves from HOG and SIFT edge/feature heuristics in the 1970s–2000s to learnable CNN learnable parameters and now Vision Transformers with positional embeddings; practical claims: sliding convolutions enable scale- and position-invariance for tumor detection, vector embeddings + nearest-neighbor search handle catalog matching, and pre-trained VLMs (demoed on Hugging Face) can caption and segment buildings but fail to segment specific retail cans or count niche boxes; fine-tuning and annotation pipelines plus deployment (about 25% of startups focus here) fix many domain gaps.
AI-written summary · about
chapters (9)
- 0:00Introduction to computer vision
- 1:55Early computer vision algorithms
- 4:21Rise of convolutional neural networks
- 6:07Few-shot learning in vision
- 7:31Vision Transformers explained
- 9:16Multimodal vision language models
- 10:40Limitations of current models
- 11:42Fine-tuning and deployment
- 12:45Future of computer vision
related talks
33mSeeing Through the Lens of Transformers: A New Era in Computer Vision
42mState of the Art of Innovation in Computer Vision
26mComputer Vision in the Era of Deep Learning
04mUsing Computer Vision to Understand Big Visual Data
35mBeyond painting buttons: computer vision techniques
44mComputer Vision in Practice
17mVision AI in 2025
20mComputer Vision: From Facial Recognition to Image Analysis