Section: New Results

Multi-dimensional indexing and clustering

Improved NV-tree

Participant : Laurent Amsaleg.

This is a joint work with Björn Þór Jónsson from the School of Computer Science, Reykjavik University, Iceland and with Herwig Lejsek, Videntifier Technologies, Iceland.

We have further improved the NV-Tree (Nearest Vector Tree) indexing techniques. It addresses the specific, yet important, problem of efficiently and effectively finding the approximate $k$-nearest neighbors within a collection of a few billion high-dimensional data points. The NV-Tree is a very compact index, as only six bytes are kept in the index for each high-dimensional descriptor. It thus scales extremely well when indexing large collections of high-dimensional descriptors. The NV-Tree efficiently produces results of good quality, even at such a large scale that the indices cannot be kept entirely in main memory any more. We have demonstrated this with extensive experiments using a collection of 2.5 billion SIFT (Scale Invariant Feature Transform) descriptors. Additional experiments involving more than 30 billion SIFT descriptors show results are still of a good quality and that disks are handled as efficiently as they can be.

Indexation of time series

Participants : Laurent Amsaleg, Romain Tavenard.

Dynamic Time Warping (DTW) is the most popular approach for evaluating the similarity of time series, but its computation is costly. Therefore, simple functions lower bounding DTW distances have been designed, accelerating searches by quickly pruning sequences that could not possibly be best matches. The tighter the bounds, the more they prune and the better the performance. Designing new functions that are even tighter is difficult because their computation is likely to become complex, canceling the benefits of their pruning. It is possible, however, to design simple functions with a higher pruning power by relaxing the no false dismissal assumption, resulting in approximate lower bound functions. We have discover how very popular approaches accelerating DTW such as LB_Keogh and LB_PAA can be made more efficient via approximations. The accuracy of approximations can be tuned, ranging from no false dismissal to potential losses when aggressively set for great response time savings. At very large scale, indexing time series is mandatory. These approximate lower bound functions can be used with $i$SAX. Furthermore, we have also observed that a $k$-means-based quantization step for $i$SAX gives significant performance gains.

Improved image indexing with asymmetric Hamming embedding

Participants : Patrick Gros, Mihir Jain, Hervé Jégou.

We have proposed [28] an improved asymmetric Hamming Embedding scheme for large scale image search based on local descriptors. The comparison of two descriptors relies on an vector-to-binary code comparison, which limits the quantization error associated with the query compared with the original Hamming Embedding method. The approach is used in combination with an inverted file structure that offers high efficiency, comparable to that of a regular bag-of-features retrieval systems, and consistently improves the search quality over the symmetric version on the two datasets used for the evaluation.

Compression techniques for nearest neighbor search

Participants : Laurent Amsaleg, Teddy Furon, Hervé Jégou, Romain Tavenard.

Part of this work on this topic was done in cooperation with Matthijs Douze and Cordelia Schmid (INRIA/Lear ).

Re-ranking with source coding

An extension of our previous work on source coding techniques for high-dimensional indexing has been proposed [29] . The goal is to index a large set of vectors, as large as 1 billion vectors, with limited CPU and memory usage. Based on the product quantization-based indexing technique [18] , we show that it is interesting to add an additional level of processing to refine the estimated distances. It consists in quantizing the difference vector between a point and the corresponding centroid. When combined with an inverted file, this gives three levels of quantization. Experiments performed on SIFT and GIST image descriptors show excellent search accuracy outperforming three state-of-the-art approaches. Compared with the original work [18] , the proposed re-ranking technique is shown to obtain better trade-off with respect to memory, efficiency and search quality.

Anti-sparse coding for approximate nearest neighbor search

Following recent works on Hamming Embedding techniques, we propose [67] a binarization method that aim at addressing the problem of nearest neighbor search for the Euclidean metric by mapping the original vectors into binary vectors ones, which are compact in memory, and for which the distance computation is more efficient.

Our method is based on the recent concept of anti-sparse coding, which exhibits here excellent performance for approximate nearest neighbor search. Unlike other binarization schemes, this framework allows, up to a scaling factor, the explicit reconstruction from the binary representation of the original vector. We also show that random projections which are used in Locality Sensitive Hashing algorithms, are significantly outperformed by regular frames for both synthetic and real data if the number of bits exceeds the vector dimensionality, i.e., when high precision is required.

Architecture-aware indexing techniques for solid state disks

Participants : Laurent Amsaleg, Gylfi Gudmundsson.

This is a joint work with Björn Þór Jónsson from the School of Computer Science, Reykjavik University, Iceland.

The scale of multimedia data collections is expanding at a very fast rate. In order to cope with this growth, the high-dimensional indexing methods used for content-based multimedia retrieval must adapt gracefully to secondary storage. Recent progress in storage technology, however, means that algorithm designers must now cope with a spectrum of secondary storage solutions, ranging from traditional magnetic hard drives to state-of-the-art solid state disks. We have analyzed the impact of storage technology on a simple, prototypical high-dimensional indexing method for large scale query processing. We found that while the algorithm implementation deeply impacts the performance of the indexing method, the setup of the underlying storage technology is equally important.