Hierarchical Matching with Side Information for Image Classification CVPR 2012 Hierarchical Matching with Side Information for Image Classification Authors Qiang Chen et al. Presenter Hyung-seok Lee (이형석)
Motivation (1/2) Bag-of-Words (BoW) Frequency histogram of visual words Not emphasize any elements with regard to image layout
Motivation (2/2) Spatial Pyramid Matching (SPM) Partitioning the image plane into fine sub-regions Not optimum for object-centered recognition problem Beyond Bags of Features : Spatial Pyramid Matching for Recognizing Natural Scene Categories, Svetlana Lazebnik et al.
Main Idea Generalized Hierarchical Matching (GHM) Extends the popular matching work Integrate prior knowledge for enhancing feature matching side information
Overview Image Local features & side information Hierarchical clustering Hierarchical structure representation Matching over each cluster
Side Information (1/4) Object Confidence Map Object position should be extremely beneficial Shape model / Appearance model
Side Information (2/4) Shape-based object detection : HOG features Appearance-based object detection : BoW features Object Confidence Map
Side Information (3/4) Visual Saliency Map Aim to detect the important content in images Explain human visual search strategies SUN : A Bayesian Framework for Saliency Using Natural Statistics, Lingyun Zhang et el.
Side Information (4/4) Use the saliency model SUN (Saliency Using Natural statistics) Based on natural image statistics
Side Information Combination Combine object confidence map & visual saliency map Design a 3 level hierarchical structure - Object confidence map is used in level 2 - Background area will be further utilized in level 3
Encoding & Matching (1/2)
Encoding & Matching (2/2) Encoding is operated on each cluster Matching over each corresponding cluster
Experiments (1/2) Caltech-UCSD Birds 200 & Oxford Flowers 17 and 102
Experiments (2/2) VOC 2007 VOC 2010
Conclusion Propose the Generalized Hierarchical Matching framework - Extends the popular pyramid matching Two novel kinds of side information are introduced to enhance object-oriented image classification tasks Object confidence map Visual saliency amp