Exemplar Metric Learning For Object Detection

Object detection is one of the most fundamental and challenging topic in computer vi-sion research area. Object detection is closely related with many research areas in computervision. Object detection has wide applications, including content-based image retrieval,video surveillance, object tracking, natural human-computer interface and intelligent trans-port system. Object detection is a very hard task, due to large intra-class variations. Partic-ularly in real application systems, the imaging conditions may be extraordinarily complex.The performance of detection algorithms will drop as a result of view difference, illumi-nation effect, object deformation and occlusion, etc. In this paper, the Exemplar MetricLearning method was proposed to handle large intra-class variations in object detection.The main content of this paper are as follows:Firstly, the author introduced research backgrounds, basic problems and research ap-proaches on object detection, summarized recent advances about object detection. In addi-tion, backgrounds of metric learning and exemplar-based methods were also briefly intro-duced.Secondly, a brief review on metric learning methods and recent advances was given.Relevant applications of metric learning in computer vision area were introduced, includingcontent-based image retrieval, image classification and face verification.Thirdly, the author proposed an exemplar metric learning approach for object detec-tion. The proposed method well utilized advantages of metric learning on multi-directiondiscriminant analysis. This approach can well enhance the performance of object detection.Then, the proposed algorithm was tested on two object categories including side-viewcar and human face. Experimental results on UIUC-Car and FDDB proved that this al-gorithm is effective in handling intra-class variations and complex backgrounds on objectdetection.A complete conclusion and possible future works are given at the end.