Bird Species Categorization Using Pose Normalized Deep Convolutional Nets

• DOI: 10.48550/arxiv.1406.2952
• Publication Date: 2014-01-01
• Authors: Steve Branson; Grant Van Horn; Serge Belongie; Pietro Perona
• Keywords: FOS: Computer and information sciences; Computer Vision and Pattern Recognition (cs.CV); Computer Science - Computer Vision and Pattern Recognition; 0202 electrical engineering, electronic engineering, information engineering; 02 engineering and technology; 004

We propose an architecture for fine-grained visual categorization that approaches expert human performance in the classification of bird species. Our first computes estimate object's pose; this is used to compute local image features which are, turn, classification. The are computed by applying deep convolutional nets patches located and normalized pose. perform empirical study a number pose normalization schemes, including investigation higher order geometric warping functions. novel graph-based clustering algorithm learning compact space. detailed state-of-the-art feature implementations fine-tuning observe model integrates lower-level layers with pose-normalized extraction routines higher-level unaligned works best. experiments advance on species recognition, large improvement correct rates over previous methods (75% vs. 55-65%).