Computer vision and image processing enable machines to interpret visual data, evolving from early edge detection to deep learning models like CNNs and segmentation frameworks. Key milestones include the Sobel filter, LeNet-5, AlexNet's ImageNet victory, YOLO, and the Segment Anything Model.
Aristotle observes the principle of the camera obscura, where light passing through a small hole projects an inverted image. This early understanding of optics lays groundwork for image formation. #history #optics
Ibn al-Haytham (Alhazen) publishes his influential work on optics, explaining vision, light, and the camera obscura. His scientific method influences later image understanding. #science #history
Johann Zahn describes a portable camera obscura that could be used for drawing, anticipating photographic cameras. This advances image capture technology. #history #technology
Joseph Nicéphore Niépce creates the first permanent photograph using a camera obscura and bitumen-coated pewter plate. This marks the birth of photography. #photography #history
Louis Daguerre's daguerreotype process is announced to the public, producing detailed images on silver-plated copper. It becomes the first widely used photographic method. #photography #history
George Eastman launches the Kodak camera with roll film, making photography accessible to amateurs. The slogan 'You press the button, we do the rest' popularizes snapshot photography. #photography #technology
The Dartmouth Conference coins the term 'artificial intelligence' and includes early discussions on computer vision. This event marks the formal birth of AI as a field. #AI #history
David Hubel and Torsten Wiesel discover that neurons in the cat visual cortex respond to oriented edges, inspiring later computer vision models. Their work earns a Nobel Prize. #neuroscience #vision
Larry Roberts' MIT PhD thesis describes extracting 3D information from 2D line drawings, one of the earliest computer vision systems. It introduces edge detection and object recognition. #computervision #history )
Marvin Minsky and Seymour Papert launch the MIT Summer Vision Project, aiming to build a complete vision system. Though overly ambitious, it spurs foundational research in computer vision. #AI #computervision
Irwin Sobel and Gary Feldman present the Sobel operator at Stanford, an edge detection algorithm using 3x3 convolution kernels. It becomes a fundamental tool in image processing. #imageprocessing #computervision
John Canny develops the Canny edge detector, a multi-stage algorithm for optimal edge detection. It remains widely used in computer vision. #computervision #imageprocessing
The fast Fourier transform (FFT) is applied to image processing, enabling efficient frequency-domain analysis. This technique is crucial for filtering and compression. #imageprocessing #math
David Marr publishes 'Vision', proposing a computational framework for visual perception. His hierarchical approach influences computer vision research. #computervision #neuroscience )
Kunihiko Fukushima introduces the Neocognitron, a neural network model for pattern recognition inspired by the visual cortex. It is a precursor to convolutional neural networks. #deeplearning #computervision
David Hubel and Torsten Wiesel receive the Nobel Prize for their discoveries on information processing in the visual system. Their work continues to inspire computer vision architectures. #neuroscience #vision
The Joint Photographic Experts Group (JPEG) standard is introduced for image compression, using discrete cosine transform. It becomes the most widely used image format. #imageprocessing #technology
Yann LeCun publishes a paper on convolutional neural networks (CNNs) for handwritten digit recognition, laying the foundation for modern deep learning in vision. #deeplearning #computervision
Yann LeCun et al. develop LeNet-5, a CNN that achieves high accuracy on MNIST handwritten digit recognition. It demonstrates the power of CNNs for image classification. #deeplearning #computervision
David Lowe introduces the Scale-Invariant Feature Transform (SIFT), a robust local feature descriptor for object recognition. SIFT becomes a cornerstone of feature-based computer vision. #computervision #imageprocessing
Paul Viola and Michael Jones develop a real-time face detection algorithm using Haar-like features and AdaBoost. It enables efficient face detection in images and video. #computervision #facedetection
Navneet Dalal and Bill Triggs introduce Histogram of Oriented Gradients (HOG) for pedestrian detection. HOG features become widely used in object detection. #computervision #objectdetection
Fei-Fei Li initiates the ImageNet project, a large-scale dataset of labeled images. ImageNet becomes a benchmark for image classification and drives deep learning advances. #computervision #dataset
The ImageNet dataset is publicly released, containing over 14 million labeled images across 20,000 categories. It catalyzes competition and progress in computer vision. #computervision #dataset
Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton's AlexNet achieves a top-5 error rate of 15.3% in the ImageNet Large Scale Visual Recognition Challenge, significantly outperforming traditional methods. This breakthrough popularizes deep CNNs. #deeplearning #computervision
OverFeat introduces a multi-scale CNN for object detection and classification, demonstrating the power of sliding-window approaches. It influences later detection models. #computervision #objectdetection
Ian Goodfellow proposes GANs, where a generator and discriminator compete to produce realistic images. GANs revolutionize image generation and manipulation. #deeplearning #imagegeneration
VGGNet shows the importance of depth with small filters, while GoogLeNet introduces the Inception module for efficient computation. Both achieve top results in ImageNet. #deeplearning #computervision
Olaf Ronneberger et al. propose U-Net, a CNN architecture for biomedical image segmentation with skip connections. It becomes widely adopted in medical imaging. #computervision #segmentation
Kaiming He et al. introduce Residual Networks (ResNet) with skip connections, enabling training of very deep networks (152 layers). ResNet wins ImageNet and becomes a foundational architecture. #deeplearning #computervision
Joseph Redmon et al. introduce You Only Look Once (YOLO), a single-shot object detection system that predicts bounding boxes and class probabilities in one pass. YOLO achieves real-time performance. #computervision #objectdetection )
DeepMind's AlphaGo beats world champion Lee Sedol at Go, using deep reinforcement learning and computer vision techniques. It demonstrates AI's ability to master complex visual tasks. #AI #deeplearning
Vaswani et al. propose the Transformer model for sequence tasks, using self-attention. It later inspires Vision Transformers (ViT) for image classification. #deeplearning #AI )
Kaiming He et al. introduce Mask R-CNN, extending Faster R-CNN with a mask branch for pixel-level instance segmentation. It sets new standards in segmentation. #computervision #segmentation
DeepMind's BigGAN achieves state-of-the-art image generation on ImageNet, producing high-resolution and diverse images. It advances generative models. #imagegeneration #deeplearning
Google AI releases BERT, a transformer-based model for NLP. While not vision-specific, BERT's self-supervised learning influences vision-language models. #NLP #deeplearning )
Dosovitskiy et al. show that a pure transformer applied directly to image patches can outperform CNNs on image classification. ViT becomes a popular alternative to CNNs. #computervision #deeplearning
OpenAI releases DALL-E, a transformer model that generates images from textual descriptions. It demonstrates the power of multimodal AI. #AI #imagegeneration
Stability AI releases Stable Diffusion, an open-source text-to-image model using latent diffusion. It democratizes AI image generation and spurs creative applications. #AI #imagegeneration
Meta AI releases the Segment Anything Model (SAM), a promptable segmentation system trained on a massive dataset (SA-1B). SAM enables zero-shot segmentation of any object. #computervision #segmentation
OpenAI's GPT-4V (Vision) extends GPT-4 to accept image inputs, enabling multimodal reasoning. It marks a step toward general visual understanding. #AI #multimodal
OpenAI unveils Sora, a diffusion model that generates realistic videos from text prompts. It pushes the boundaries of video generation and understanding. #AI #videogeneration )