Visual concepts and compositional voting

• DOI: 10.4310/amsa.2018.v3.n1.a5
• Publication Date: 2018-03-27T18:36:13Z
• Authors: Jianyu Wang; Zhishuai Zhang; Cihang Xie; Yuyin Zhou; Vittal Premachandran; Jun Zhu; Lingxi Xie; Alan Yuille
• 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; 006; 02 engineering and technology

It is very attractive to formulate vision in terms of pattern theory \cite{Mumford2010pattern}, where patterns are defined hierarchically by compositions elementary building blocks. But applying real world images currently less successful than discriminative methods such as deep networks. Deep networks, however, black-boxes which hard interpret and can easily be fooled adding occluding objects. natural wonder whether better understanding networks we extract blocks used develop theoretic models. This motivates us study the internal representations a network using vehicle from PASCAL3D+ dataset. We use clustering algorithms population activities features set visual concepts show visually tight correspond semantic parts vehicles. To analyze this annotate these vehicles their create new dataset, VehicleSemanticParts, evaluate unsupervised part detectors. that perform fairly well but outperformed supervised Support Vector Machines (SVM). next give more detailed analysis how they relate parts. Following this, for simple theoretical model, call compositional voting. In model several combine detect approach significantly like SVM trained specifically detection. Finally, return studying occlusion creating an annotated dataset with occlusion, called VehicleOcclusion, voting outperforms even when amount becomes large.