Course description

Billions of images are on the web---how can you find the ones you are interested in?  How could photo collections on social media be indexed automatically by the people or events they contain?  How could we interact with a computer using natural gestures or facial expressions?  How can a robot identify objects in complex environments, or navigate uncharted territory?  After capturing video with a wearable camera for days on end, how to determine those snapshots worth keeping?  How can we develop augmented reality systems that overlay visualizations relevant to the real-world content in sight, e.g., a menu for the restaurant you just passed on the street, or a field guide entry for the unusual insect you encountered while hiking?

All such questions demand high-level computer vision.  In computer vision, the goal is to develop methods that enable a machine to “understand” or analyze images and videos.   In this introductory vision course, we will explore fundamental topics in the field ranging from low-level feature extraction to high-level visual recognition.  Throughout, we will emphasize machine learning-based methods.  While we will motivate the concepts from the vision problems, the learning algorithms we will study are also useful tools for other domains in AI and beyond.  Note, due to our emphasis on learning methods (and the time constraints for the syllabus), this course will omit or treat only briefly some core aspects of computer vision, such as multi-view geometry, 3d reconstruction, and tracking.

This course is intended for honors upper-level undergraduate students. 


Syllabus

Details on prerequisites, course requirements, textbooks, and grading policy are here.    A high-level summary of the syllabus is as follows:

I. Features and filters: low-level vision

• Linear filters
• Edges and contours
• Binary image analysis
• Background subtraction
• Texture
• Motion and optical flow

II. Grouping and fitting: mid-level vision

• Segmentation and clustering algorithms
• Hough transform
• Fitting lines and curves
  
• Robust fitting, RANSAC
• Deformable contours
• Interactive segmentation

III. Multiple views

• Local invariant feature detection and description
• Image transformations and alignment
• Planar homography
• Epipolar geometry and stereo
• Object instance recognition
  

IV.  Recognition: high-level vision

• Object/scene/activity categorization
• Object detection
  
• Supervised classification algorithms
  
• Probabilistic models for sequence data
• Learning to rank
• Active learning
• Dimensionality reduction and manifold learning
  
• Non-parametric methods and big data
  
• Deep learning, convolutional neural networks
• Crowdsourcing and dataset creation