摘要
Object recognition, which consists of classification and detection, has two important attributes for robustness： 1） closeness： detection windows should be as close to object locations as possible, and 2） adaptiveness： object matching should be adaptive to object variations within an object class. It is difficult to satisfy both attributes using traditional methods which consider classification and detection separately; thus recent studies propose to combine them based on confidence contextualization and foreground modeling. However, these combinations neglect feature saliency and object structure, and biological evidence suggests that the feature saliency and object structure can be important in guiding the recognition from low level to high level. In fact, object recognition originates in the mechanism of ＂what＂ and ＂where＂ pathways in human visual systems. More importantly, these pathways have feedback to each other and exchange useful information, which may improve closeness and adaptiveness. Inspired by the visual feedback, we propose a robust object recognition framework by designing a computational visual feedback model （VFM） between classification and detection. In the ＂what＂ feedback, the feature saliency from classification is exploited to rectify detection windows for better closeness; while in the ＂where＂ feedback, object parts from detection are used to match object structure for better adaptiveness. Experimental results show that the ＂what＂ and ＂where＂ feedback is effective to improve closeness and adaptiveness for object recognition, and encouraging improvements are obtained on the challenging PASCAL VOC 2007 dataset.
Object recognition, which consists of classification and detection, has two important attributes for robustness： 1） closeness： detection windows should be as close to object locations as possible, and 2） adaptiveness： object matching should be adaptive to object variations within an object class. It is difficult to satisfy both attributes using traditional methods which consider classification and detection separately; thus recent studies propose to combine them based on confidence contextualization and foreground modeling. However, these combinations neglect feature saliency and object structure, and biological evidence suggests that the feature saliency and object structure can be important in guiding the recognition from low level to high level. In fact, object recognition originates in the mechanism of ＂what＂ and ＂where＂ pathways in human visual systems. More importantly, these pathways have feedback to each other and exchange useful information, which may improve closeness and adaptiveness. Inspired by the visual feedback, we propose a robust object recognition framework by designing a computational visual feedback model （VFM） between classification and detection. In the ＂what＂ feedback, the feature saliency from classification is exploited to rectify detection windows for better closeness; while in the ＂where＂ feedback, object parts from detection are used to match object structure for better adaptiveness. Experimental results show that the ＂what＂ and ＂where＂ feedback is effective to improve closeness and adaptiveness for object recognition, and encouraging improvements are obtained on the challenging PASCAL VOC 2007 dataset.
基金
This work was supported by the National Basic Research 973 Program of China under Grant No. 2012CB316302, the National Natural Science Foundation of China under Grant Nos. 61322209 and 61175007, the National Key Technology Research and Development Program of China under Grant No. 2012BAH07B01.Thank Steve Maybank for the revision.
