One of the tasks under ALADDIN is that of video event detection. Sparse projection has been shown to be highly effective in related domains, e.g., image denoising and scene / object classification. However, practical application to large scale problems such as video analysis requires efficient versions of sparse projection algorithms such as Orthogonal Matching Pursuit (OMP). In particular, random projection based locality sensitive hashing (LSH) has been proposed for OMP. We present a novel technique called Comparison Hadamard random projection (CHRP) for further improving the efficiency of LSH within OMP. CHRP combines two techniques: (1) The Fast Johnson-Lindenstrauss Transform (FJLT) which uses a randomized Hadamard transform and sparse projection matrix for LSH, and (2) Achlioptas' random projection that uses only addition and comparison operations. Our approach provides the robustness of FJLT while completely avoiding multiplications, and is demonstrated for image denoising, scene classification, and video categorization.
Host: Ashok Veeraraghavan
Biography of Shiv N. Vitaladevuni:
Shiv Vitaladevuni is a Scientist at the Speech, Language and Multimedia Unit at Raytheon BBN Technologies. His research interests include semantic analysis of images and videos, machine learning and optimization. Dr. Vitaladevuni received his Ph.D. from University of Maryland under Prof. Larry Davis, specializing in video-based action recognition. Prior to joining BBN, he worked for 3 years at Howard Hughes Medical Institute (HHMI) building an end-to-end system for reconstructing neuronal connectivity from electron micrographs. Dr. Vitaladevuni is the Co-PI in an IARPA program in which he leads the research in discovering theories for financial market behavior from heterogeneous data. He is a contributor to several ongoing programs at BBN including, IARPA ALADDIN for web video analysis, DARPA MADCAT for document image analysis, and DARPA DCAPS for detecting subtle psychological distress indicators from web text and EEG.