Object Classification and Detection Based on Deep Representation of Object Windows
Abstract: Object classification and object detection are important problems in computer vision and pattern recognition. The most popular methodology in object classification and detection is estimating object information from whole windows (windows that contain a whole object), such as the bag-of-visual-words model and the deformable part-based model. This methodology ignores the value of estimating object information from part windows (windows that contain meaningful parts of an object), and thus leads to insufficiency in representing object windows and describing the relationship between objects and their context. Accordingly, it is an important chance to understand the insight of object windows as well as their effective representation. In this proposal, we propose to study the object window, including its physical meaning and mathematical models. Specifically, we propose to model object windows based on deep representation, which is then applied to constructing a unified framework for object classification and detection. Also, we elaborate its necessity, value and feasibility. This project is crucial to bring technical innovation for object classification and detection. We believe that this work has the potential to become an influential study and create a new direction in the field of object classification and detection.
Keywords: computer vision; pattern recognition; deep learning; object classification; object detection
Contact:
HUANG Yongzhen
E-mail: yzhuang@nlpr.ia.ac.cn