loading page

Ternary and Binary Quantization for Efficient Exemplar-based Classification
  • +1
  • weizhi lu ,
  • Mingrui Chen ,
  • kai guo ,
  • Weiyu Li
weizhi lu
Shandong University, Shandong University, Shandong University

Corresponding Author:[email protected]

Author Profile
Mingrui Chen
Author Profile


In the work, we study the exemplar-based classification method, which categorizes novel objects by comparing their similarity to the exemplars previously stored by class. The exemplar storage and similarity calculation will pose great challenges for large-scale classification. To reduce data complexity, we propose to first reduce data dimension by random projection and then quantize the projections to binary or ternary codes. Intuitively, the extreme quantization will degrade the accuracy of classification due to significant information loss. Interestingly, however, we observe that the quantization could provide comparable or even superior accuracy, if the data to be quantized are the sparse features generated with common filters. Furthermore, this quantization property could be preserved in the random projections of sparse features, if both the features and random projection matrices are sufficiently sparse. By conducting extensive experiments, we validate and analyze this intriguing property.