Large Scale Image Classification: Fast Feature Extraction, Multi-codebook Approach and Multi-core SVM Training

Abstract

The usual frameworks for image classification involve three steps: extracting features, building codebook and encoding features, and training the classifier with a standard classification algorithm (e.g. SVMs). However, the task complexity becomes very large when applying these frameworks on a large scale dataset like ImageNet containing more than 14million images and 21,000 classes. The complexity is both about the time needed to perform each task and the memory and disk usage (e.g. 11TB are needed to store SIFT descriptors computed on the full dataset). We have developed a parallel version of LIBSVM to deal with very large datasets in reasonable time. Furthermore, a lot of information is lost when performing the quantization step and the obtained bag-of-words (or bag-of-visual-words) are often not enough discriminative for large scale image classification. We present a novel approach using several local descriptors simultaneously to try to improve the classification accuracy on large scale image datasets.We show our first results on a dataset made of the ten largest classes (24,807 images) from ImageNet.KeywordsImage RepresentationLinear KernelLarge Scale DatasetSift DescriptorLarge Scale ImageThese keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.