Visual event recognition

Abstract

This report summarizes the work that has been done in the final year project of recognizing visual events in videos. It starts with image recognition, in which im- ages are represented in spatial pyramids. Such representations are then input into SVM and KNN for recognition. In video recognition, bag of words and special- ized Gaussian Mixture Models are employed to represent videos, and respective distance calculation is used to measure video-to-video distance. These distance matrices are then input into SVM for recognition using different kernel types. Also, four domain adaptation methods are implemented to recognize Kodak con- sumer videos using Youtube videos. Adaptive multiple kernel learning achieves the best and improves the mean average precision from 44.33% to 61.40%. Last but not least, a web-based demo system is implemented in two modes to visually demonstrate the underlying recognition system.