Motivation Moving from lab-controlled saliency detection to real world application requires more than “better AUC numbers” Comparing different saliency detection algorithms requires a mechanism to evaluate uncertainty and statistical significance To be applicable to wide range of videos, estimation needs to be HVS-inspired rather than data-fitted, hence unsupervised. UNSUPERVISED UNCERTAINTY ANALYSIS FOR VIDEO SALIENCY DETECTION Tariq Alshawi, Zhiling Long and Ghassan AlRegib {talshawi, zhiling.long, alregib}@gatech.edu School of Electrical and Computer Engineering, Georgia Institute of Technology, Atlanta, GA 30332, USA 1. Chenlei Guo; Liming Zhang, "A Novel Multiresolution Spatiotemporal Saliency Detection Model and Its Applications in Image and Video Compression," in Image Processing, IEEE Transactions on , vol.19, no.1, pp.185-198, Jan. 2010 2. F. W. M. Stentiford, “Attention based Auto Image Cropping,” Workshop on Computational Attention and Applications, ICVS, Bielefeld, March 21-24, 2007. 3. K. Debattista, L.P. Santos, A. Chalmers, Accelerating the irradiance cache through parallel component-based rendering, in 6th Eurographics Symposium on Parallel Graphics Visualization. Eurographics, May 2006, pp. 27-34. Uncertainty Framework Proposed Method Experiments Data Public CRCNS database 50 video clips (640 x 480), 5-90 seconds each, 30 frames/sec Street scenes, TV programs, video games, etc. Ground truth eye fixation data from human subjects under freeview condition Feature Extraction Feature Extraction Localized 3D FFT Spectral Decomposition Center-Surround Comparison F F t F s Center-Surround Comparison + S t S s S E t E s f t f x f y ᶱ M (a 0 , b 0 , c 0 ) a 0 b 0 c 0 N (a 0 , b 0 , 0) O Auto-Cropping 2 Rendering 3 Video Processing Algorithm Saliency Detection Algorithm Decision Making Module Risk Assessment Module Uncertainty Estimation Input Video Saliency Map Uncertainty Map Decision Map Output Video Cost Compression 1 Evaluation Methodology Input Video Saliency Detection Evaluation Dataset Results Expanded Eye- fixation map Saliency Map True Uncertainty - Uncertainty Estimation Estimated Uncertainty Fixed Threshold Receiver Operation Characteristics (ROC) M N D Saliency Map S Spatial Neighbors Frame# N Frame# N+1 Frame# N–1 Pixel of Interest Temporal Neighbors 0.6 1 0.6 1 -0.75 1 0.6 1 0.6 Estimated Uncertainty U e True Uncertainty U tr for beverly05, frame 5 Saliency value at pixel x Average value around pixel x