Introduction to Computer Vision Lecture 1

Computer Vision Lecture Notes

• Definition: Computer vision is a process that produces a useful description from images of the external world, avoiding irrelevant information. It constructs explicit and meaningful descriptions of physical objects. Its goal is to allow for useful decisions about physical objects and scenes based on sensed images.

Introduction to Computer Vision

• Lecture One: This is the introduction lecture.

Course Etiquette

• Be Quiet: Please be quiet during the lecture.

Related Fields – Image Processing

• Image Processing: Involves processing images, typically using low-level techniques like compression and edge detection. It focuses on quantitative measurements.
• Computer Vision: Goes beyond image processing. It involves extracting symbolic descriptions and using higher-level techniques like object recognition. Semantic (quantitative and qualitative) outputs are often associated.
• Relationship: Image processing is often used in computer vision, though the fields overlap and definitions aren't always clear cut.

Related Fields – Pattern Recognition

• Pattern Recognition: Recognition of patterns, often represented as feature vectors. Techniques are useful in 2D and constrained 3D image recognition but limited for general 3D problems.

Related Fields – Photogrammetry

• Photogrammetry: Accurately measures properties from images. Historically focused on remote sensing, especially from airplanes and satellites. It shares similar techniques with computer vision.

Related Fields – Computer Graphics

• Inverse Relationship: Computer vision is the inverse of computer graphics. Computer graphics creates 3D models from information like object locations, lighting, and parameters.
• Unique Processes: The forward process (graphics) is unique, but the inverse process (vision) isn't.

The Summer Vision Project

• Purpose: The project sought to effectively use summer workers for a significant part of the visual system. The task was designed to be segmented into smaller sub-problems that allowed for independent work and teamwork in understanding "pattern recognition".

Low-Level Vision

• Resizing: This includes different techniques like Nearest Neighbor and Bilinear resizing.
• Adjustments: This includes operations like grayscale conversion, image adjustments regarding exposure, saturation, hue and edges.
• Segmentation: The process includes color segmentation using the image capture device.

Mid-Level Vision

• Panorama Stitching: Creates a wider view by stitching together multiple images, aligning them carefully.
• Multi-View Stereo: Creates 3D information from multiple images with different viewpoints.
• Structured Light: Projecting a pattern of light to a surface helps create a 3D model of the structure.
• Range Finding: Determines the distance to objects in the scene based on the images or through using devices.
• Optical Flow: Identifies changes in images or videos based on motion, typically using tracked features in each image.
• Time Lapse: Combines various images with the time stamps to determine how a scene changes over time (e.g., a 360 camera on the space needle).

High-Level Vision

• Classification: Figure out what is in the image.
• Tagging: Identify and describe all objects and elements in the image, potentially adding details.
• Detection: Locating and identifying all specific elements in an image and their specific positions.
• Semantic Segmentation: Dividing an image into different areas according to the class (semantics) of the image objects. This typically labels different aspects of the scene, like "roadway," "car," "vegetation," and "building."
• Instance Segmentation: Further categorizes objects, so if there are multiple instances of the same class, it can uniquely identify each (such as different cars in traffic).
• Additional High-Level Tasks: Single image 3D data, game-playing, superresolution, and image retrieval.

Conclusion

• Thank You: An expression of gratitude to those who attend the lecture.