Japanese / English

Detail of Publication

Text Language Japanese
Authors Kazuto NOGUCHI, Koichi KISE, Masakazu IWAMURA
Title Experimental Evaluations of a Memory Reduction Methodfor Object Recognition Based on Local Descriptors
Journal Proceedings of MIRU 2008
Presentation number OS10-3
Pages pp.251-258
Reviewed or not Not reviewed
Month & Year July 2008
Abstract For object recognition based on local descriptors, balancing the recognition rate, processing time and memory requirement is an important issue. In conventional methods, local descriptors are vector-quantized to “visual words” for describing images. Although many visual words allow us to improve the recognition rate for recognizing instances (specific objects), they also pose a problem of processing time and memory requirement. On the other hand, even if a recognition method employs local descriptors without vector quantization, approximate nearest neighbor search enables us to cut the processing time drastically. In this report, we propose a method of memory reduction applicable even to methods without vector quantization. The proposed method reduces the memory requirement by limiting the number of bits for the record of each dimension of feature vectors. From experimental results on 100,000 images, it has been confirmed that a representation with 2 bit / dimension subtly decreases the recognition rate, but hardly changes the processing time.
Back to list