High-dimensional signature compression for large-scale image classification

摘要: We address image classification on a large-scale, i.e. when large number of images and classes are involved. First, we study accuracy as function the signature dimensionality training set size. show experimentally that larger set, higher impact accuracy. In other words, high-dimensional signatures important to obtain state-of-the-art results datasets. Second, tackle problem data compression very (on order 105 dimensions) using two lossy strategies: reduction technique known hash kernel an encoding based product quantizers. explain how gain in storage can be traded against loss and/or increase CPU cost. report databases — ImageNet dataset lM Flickr showing reduce our by factor 64 128 with little Integrating decompression classifier learning yields efficient scalable algorithm. On ILSVRC2010 74.3% at top-5, which corresponds 2.5% absolute improvement respect state-of-the-art. subset 10K top-1 16.7%, relative 160%