Learning Scalable Discriminative Dictionary with Sample Relatedness

Attributes are widely used as mid-level descriptors of object properties in object recognition and retrieval. Mostly,such attributes are manually pre-defined based on domainknowledge, and their number is fixed. However, pre-definedattributes may fail to adapt to the properties of the data athand, may not necessarily be discriminative, and/or maynot generalize well. In this work, we propose a dictionarylearning framework that flexibly adapts to the complexity ofthe given data set and reliably discovers the inherent discriminative middle-level binary features in the data. Weuse sample relatedness information to improve the gener-alization of the learned dictionary. We demonstrate thatour framework is applicable to both object recognition andcomplex image retrieval tasks even with few training examples. Moreover, the learned dictionary also help classifynovel object categories. Experimental results on the Animals with Attributes, ILSVRC2010 and PASCAL VOC2007datasets indicate that using relatedness information leads tosignificant performance gains over established baselines.

Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Columbus, Ohio

Vision