History Of Computer Vision

The first known digital image scanner, as we understand it today, was the "Cyclograph" developed by Dr. Russell A. Kirsch and his team by transforming images into grids of numbers.Kirsch and his colleagues at NBS, who had developed the nation's first programmable computer, the Standards Eastern Automatic Computer (SEAC), created a rotating drum scanner and programming that allowed images to be fed into it. The first image scanned was a head-and-shoulders shot of Kirsch's three-month-old son Walden.

Resource: First Digital Image

Commonly referred to as father of computer vision Larry Roberts presented his thesis on Machine Perception of Three-Dimensional Solids at MIT on June 1963 laying one of the foundations in the field.This thesis explored methods for understanding three-dimensional objects from two-dimensional images, which is a foundational concept in computer vision.

Paper: Machine Perception of Three-Dimensional Solids 

In their thesis in IEEE computer conference in 1967 Woodrow W. Bledsoe and I. Kanter presented their research which used a combination of edge detection and feature matching to recognize human faces. It was one of the first successful computer vision systems to be developed for face recognition.

Wiki: Facial Recognition System

Resource: The Secret History of Facial Recognition

The Hough transform was developed by Richard Duda and Peter Hart. The method is widely used for detecting simple geometric shapes like lines and circles in images.The key idea behind the Hough Transform is to represent each edge point not in the image space (x, y) but in a parameter space (usually represented as ρ (rho) and θ (theta)).

Wiki: Hough Transform

Resource: USE OF THE HOUGH TRASFORMTION TO DETECT LINES AND CURVES IN PICTURES

Automatix, Inc. was a company that played a pioneering role in the field of industrial automation and computer vision. It was founded in the early 1980s and specialized in developing computer vision and robotic systems for various industrial applications. Automatix is particularly notable for its work in the integration of computer vision technology into manufacturing processes.Their system was used for tasks such as quality control and inspection in manufacturing, including the inspection of electronic components for defects.

Wiki: Automatix Inc.

The RANSAC algorithm was i introduced by Martin A. Fischler and Robert C. Bolles in their paper "Random Sample Consensus: A Paradigm for Model Fitting with Applications to Image Analysis and Automated Cartography" in 1981. The Algorithm majorly focuses on Location determination problem, to identify a point in space based on certain predetermined locations.

Wiki: Random Sample Consensus

Paper: Random Sample Consensus: A Paradigm for Model Fitting

David Marr published the book Vision. This book provides a comprehensive theoretical framework for understanding vision and has had a significant impact on the field of computer vision and visual perception.

Book: Vision 

The Deriche edge detector is an edge detection technique used in computer vision and image processing. It's based on the concept of computing gradient information in an image to locate edges or boundaries between different objects or regions. The Deriche edge detector is a recursive filter designed to find edges while suppressing noise in the image.

Paper: Using Canny's criteria to derive a recursively implemented optimal edge detector

It is a mathematical model used for image segmentation and restoration. The function is particularly useful for segmenting images into distinct regions while preserving important edges and boundaries. It is used to address the problem of piecewise constant approximation of images, which means dividing an image into regions where the intensity or color is approximately constant.

Wiki: Mumford Shah Functional

Resource: Optimal Approximations by Piecewise Smooth Functions and Associated Variational Problems

Introduced by Tony Lindeberg, it is a method for detecting blob-like structures in images at various scales. It is essential for object recognition and shape analysis.It forms the basis for various feature detection and extraction techniques used in computer vision.

Paper: Detecting salient blob-like image structures and their scales with a scale-space primal sketch

The SIFT (Scale-Invariant Feature Transform) algorithm is a computer vision and image processing technique used for detecting and describing local features in images. SIFT is known for its robustness to various image transformations, such as scaling, rotation, and changes in viewpoint, making it a valuable tool for tasks like object recognition, image stitching, and more.

Wiki: Scale-Invariant Feature Transform

Paper: Object Recognition from Local Scale-Invariant Features

Haar Cascades, also known as Haar-like features and Haar Cascade Classifiers, are a machine learning object detection method used to identify objects in images or video. They were developed by Paul Viola and Michael Jones and were introduced in the early 2000s. Haar Cascades are particularly well-known for their efficiency in detecting faces but can be used for a variety of object recognition tasks.

Wiki: Haar Like Feature   

Paper: Rapid object detection using a boosted cascade of simple features 

First introduced in 1986 by Robert K. McConnell. Later The Histogram of Oriented Gradients (HOG) algorithm was introduced by Navneet Dalal and Bill Triggs, is a feature descriptor used for object detection. It calculates histograms of gradient orientations in image regions, making it effective for tasks like human detection and object recognition.

Wiki: Histogram of oriented gradients  

Paper: Histograms of oriented gradients for human detection 

Developed by Herbert Bay, Tinne Tuytelaars, and Luc Van Gool, is a feature detection and description algorithm in computer vision.The SURF algorithm consists of interest point detection, local neighborhood description, and matching for efficient and robust feature detection and matching in computer vision.

Wiki: Speeded up robust features

Paper: SURF: Speeded Up Robust Features 

The PASCAL VOC (Visual Object Classes) and ImageNet competitions were introduced to advance object recognition in computer vision by providing large-scale datasets and benchmark challenges. These competitions indeed played a crucial role in driving advancements in computer vision algorithms, particularly in the field of object recognition and classification.

Paper: The PASCAL Visual Object Classes (VOC) Challenge

ILSVRC stands for the "ImageNet Large Scale Visual Recognition Challenge." It is an annual competition in the field of computer vision and image recognition. ILSVRC focuses on large-scale image classification and object detection tasks. mageNet was one of the earliest large-scale image datasets.The competition typically involved challenges like image classification and object detection, where participants had to develop algorithms and models to correctly classify objects within images.

Resource: ImageNet Large Scale Visual Recognition Challenge (ILSVRC)

Microsoft Common Objects in Context (COCO) dataset was indeed introduced in 2014 and has become a prominent dataset in the field of computer vision. It is known for its diversity of object categories and complex scenes, making it a valuable resource for various computer vision tasks, including object detection, image segmentation, and image captioning.

Resource: Coco Dataset 

VGG is one of the classic deep learning architectures that has been widely used in computer vision tasks, particularly in image classification. Although it is not the most modern architecture, it is still used for educational purposes and as a baseline for various vision-related tasks. The simplicity of the architecture and its effectiveness make it a useful reference in computer vision.

Paper: Very Deep Convolutional Networks for Large-Scale Image Recognition

Deepdream was introduced by google created by Google engineer Alexander Mordvintsev, it is a convolutional neural network method to enhance and modify images in a unique and psychedelic way. DeepDream operates iteratively, enhancing and exaggerating the patterns or features that the neural network recognizes. This process is called "dreaming." The deeper layers of the neural network tend to recognize more complex and abstract features, resulting in visually surreal and intricate images.

Wiki: DeepDream  

Resource: DEEP DREAM GENERATOR 

Is a deep learning-based object detection framework that was introduced in 2015 by Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. It builds upon the R-CNN (Regions with Convolutional Neural Networks) family of models but significantly improves their speed and accuracy.Faster R-CNN introduced the concept of region proposal networks, which significantly improved the speed and accuracy of object detection by integrating deep learning techniques.

Wiki: Region Based Convolutional Neural Networks  

Paper: Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks 

Mask R-CNN is an extension of the Faster R-CNN object detection framework, and it was developed in 2017 by Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask R-CNN is primarily used for instance segmentation, allowing it to detect objects in images and also provide detailed pixel-level masks for each individual object, thereby segmenting them with high precision. It's a powerful tool for various computer vision tasks that require detailed object segmentation.

Paper: Mask R-CNN 

EfficientNet, on the other hand, is a more recent architecture designed to strike a balance between model accuracy and computational efficiency. It has gained popularity due to its superior performance on image classification tasks and its efficiency, which makes it suitable for various real-world applications. EfficientNet models are often used in scenarios where computational resources are limited or when efficient deployment on edge devices is required.

Paper: EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks

GPT-3, developed by OpenAI, is a large language model known for its natural language processing capabilities. While it was primarily designed for text-related tasks, it has been shown to be versatile and applicable to vision-language tasks as well.It demonstrated the versatility of large language models for vision-language tasks, allowing the generation of natural language descriptions from images.