Artificial Intelligence (AI) research, with its roots dating back to the Dartmouth Conference in 1956, has come a long way.
However, one particular facet of AI, known as computer vision (CV), which focuses on replicating human-like visual functions with computers, truly took off around 1980. Fast forward to today, and the CV field has witnessed remarkable growth, with AI developments further propelling its evolution.
The Evolution of Computer Vision
The CVPR (Computer Vision and Pattern Recognition) conference, established in 1983, has played a pivotal role in shaping the CV landscape. Over two decades, the number of papers presented at CVPR grew from about 100 to a staggering 500.
However, the real surge came in the next decade, with AI breakthroughs like deep learning causing a fourfold increase in the number of cases, surpassing 2,000. And the growth shows no signs of slowing down.
Computer vision, in essence, is about extracting physical attributes such as size, shape, color, and material from digital images. It’s the art of teaching computers to comprehend and recognize what they ‘see.’ For example, to measure an object’s length in an image, the number of pixels is calculated, and its real-world size is derived from factors like camera position and lens focal length.
Recognizing objects, like a ‘cat,’ involves creating models that describe characteristics such as triangular ears and circular eyes. The model fits the image, and the subject’s name is determined. Historically, designing these models was complex and context-dependent, limiting practicality.
The Data-Driven AI Revolution
Recent AI breakthroughs have flipped the script. Instead of crafting models based on theory, AI now leverages vast real-world data. This shift is at the heart of AI’s success.
Massive datasets, comprising images taken in diverse scenarios, serve as input, alongside ground-truth data like object names or physical measurements.
Through statistical machine learning, AI models learn the input-output relationship, enabling them to infer the correct answer for new inputs. This paradigm shift has revolutionized Computer Vision, making it more adaptable and practical.
AI’s Remarkable Feat: Depth Estimation
Consider the challenge of estimating the depth of each pixel in an image. A camera captures only two-dimensional positions and luminance values. Traditionally, techniques like stereo imaging and depth cameras were used. However, AI has rewritten the rules.
AI models, trained on extensive datasets, can now predict depth from RGB images, a task seemingly impossible through conventional methods. This transformation is a testament to AI’s ability to unlock the hidden potential in data.
Semantic Segmentation: Assigning Names to Pixels
Another AI achievement is semantic segmentation. It involves labeling each pixel in an image with the corresponding object name, like ‘table’ or ‘chair.’ This task was challenging due to the lack of a direct theory connecting pixel values to semantic labels.
AI, fueled by vast datasets, learns the intricate relationships between pixel values and object names, defying traditional limitations. This underscores AI’s capacity to bridge gaps in understanding.
Estimating Complex Object States
AI’s prowess isn’t limited to single-pixel tasks. It can estimate multidimensional data representing object states. For instance, it can measure human posture from images, analyzing joint positions and angles.
This technology has applications in health monitoring and rehabilitation. Imagine having continuous posture measurement in everyday life spaces. Cameras, strategically positioned or even attached to individuals, capture their movements. AI then deciphers these images to estimate human motion postures.
Mastering the Art of Full-Focus Images
Ever struggled to capture an image where everything is perfectly in focus, only to end up with blurry sections? AI offers a solution. It can create full-focus images from a series of shots with varying points of focus, known as a focal stack.
Traditionally, high-frequency components in local patch images were used to determine the most in-focus parts of each image in the stack. However, this method often fell short. Enter AI, equipped with the power of large-scale data and machine learning.
By generating vast datasets, including focal stack images and corresponding perfectly focused images, AI learns to discern which parts of an image should be in focus. Armed with this knowledge, AI can create sharp, in-focus images, even in scenes with complex 3D shapes.
The Rise of NeRF: Creating New Viewpoint Images
NeRF, or Neural Radiance Fields, represents a groundbreaking leap in image creation. Unlike traditional methods that relied on explicit 3D modeling, NeRF employs a neural network—an AI model—to recreate scenes.
In this revolutionary approach, the neural network takes the position as input and outputs the 3D transparency (σ) and RGB values when observed. By training on multi-viewpoint images, NeRF becomes a magician, conjuring images from entirely new perspectives. This is a quantum leap in generating images that were previously unattainable through classical techniques.
A Glimpse into the Future: Multi-Viewpoint Cameras
The future of image sensing is expanding, thanks to innovations like multi-viewpoint cameras. These devices address limitations, such as capturing images obstructed by objects like the surgeon’s head during medical procedures. By using multiple cameras simultaneously, continuous recording becomes possible, ensuring no vital moments are missed.
AI plays a pivotal role in this setup by automatically selecting unobstructed views from the array of cameras. It ensures seamless video recording and opens new possibilities for capturing perspectives that were once considered challenging.
AI-driven advancements in computer vision have revolutionized image sensing. From depth estimation to semantic segmentation and complex object state estimation, AI’s data-driven approach is transforming what was once deemed impossible into reality. This marks a new era where the power of AI unlocks the hidden potential in visual data.
