Interactive Learning and Prediction Algorithms For Computer Vision Applications.

摘要

In this dissertation, we explore three different types of interactive methods in computer vision. First, we introduce interactive variants of popular computer vision algorithms, and have applied one of them to create a practical web-based bird species recognition tool. Second, we explore a simple form of active learning that interleaves online learning and interactive labeling of structured objects, and show that it has good properties in theory and practice in terms of scalability to large datasets and complex image models, with some bounds on total annotation effort. Lastly, we investigate interactive feedback methods to researchers and annotators, with the objective of diagnosing errors due to insufficient training data, a bad model or feature space, annotation error, or insufficient computation time. We have combined these methods into a common software package and applied them to a variety of problems including object detection, pose registration and part-based methods, model sharing methods based on parts and attributes, cost-sensitive multiclass classification, and behavior detection and segmentation.