Output details

Canonical Correlation Analysis of Video Volume Tensors for Action Categorization and Detection

This innovative work casts automatic action/gesture recognition problems into a mathematically rigorous approach for feature extraction and pattern classification. It extends CCA, a standard tool to inspect linear relations of two vector sets, into a spatiotemporal domain, establishing a principled way of learning and extracting action features. The approach has reported the best accuracy on the widely used KTH benchmark. This is a key part of our action recognition work, directly leading to the prestigious junior research fellowship of Univ. of Cambridge and the industrial grant (£95k) by Samsung Electronics (hand gesture interface, Contact: Changkyu Choi, changkyu_choi@samsung.com).