Section: New Results

Feature space structuring

Random Maximum Margin Hashing

Participant : Alexis Joly.

Following the success of hashing methods for multidimensional indexing, more and more works are interested in embedding visual feature space in compact hash codes. Such approaches are not an alternative to using index structures but a complementary way to reduce both the memory usage and the distance computation cost. Several data dependent hash functions have notably been proposed to closely fit data distribution and provide better selectivity than usual random projections such as LSH. However, improvements occur only for relatively small hash code sizes up to 64 or 128 bits due to the lack of independence between the produced hash functions. In this work, we introduced a new hash function family that attempts to solve this issue in any kernel space. Rather than boosting the collision probability of close points, this method focus on data scattering. By training purely random splits of the data, regardless the closeness of the training samples, it is indeed possible to generate consistently more independent hash functions. On the other side, the use of large margin classifiers allows to maintain good generalization performances. Experiments did show that our new Random Maximum Margin Hashing scheme (RMMH) outperforms four state-of-the-art hashing methods, notably in kernel spaces. Overall, this new concept of randomly trained classifiers opens the door to many other problems including large-scale learning, visual vocabulary construction or distributed content-based retrieval methods. A paper describing RMMH was published in the proceedings of CVPR 2011 [14] .

Scalable information retrieval in distributed architectures

Participants : Mohamed Riadh Trad, Alexis Joly, Nozha Boujemaa.

Organizing media according to the occurrence of real-life events is attracting increasing interest in the multimedia community. However, whereas text based methods are now mature enough to deal with huge datasets, there are still some challenging issues managing multimedia contents. This becomes even more challenging in the context of User Generated Contents. Low-level visual metadata are indeed not simple textual or scalar values, their management requires efficient similarity search in high dimensional spaces.

Similarity search in high dimensional spaces has been the focus of many works in the database community in the recent years. State-of-the-art methods focus mainly on space partitioning techniques and more recently on hash-based probabilistic algorithms.

Although, hash-based approaches proved to be scalable, the computational cost is still too high for some real world applications and K-Nearest Neighbours Graph constructed can be more desirable than the costly online K-NN search. In fact, the basic LSH algorithm partitions the space uniformly and thus it does not exploit the clustering property of the data, which may result in slow query response and wasted space with additional hash tables. These limitations were pointed out with our scalable prototype for large scale event matching [18] .

Scaling up LSH-based techniques and applications is then closely related to buckets occupations and objects distribution within the index structure. Recent works achieve better data distribution over the buckets with guarantees on occupation. As one result, we easily bound the similarity join size and evaluate bound algorithms complexity.

Based on these works, we designed and implemented a scalable prototype for distributed similarity search and K-NN graph construction. We have made several experiments querying real world large datasets. The prototype proved to be efficient for both search and K-NN graph construction.

Ongoing experiments process a 1.2 million images dataset. Results will be submitted for publication.

Visual similarity sensitive hashing methods for semantic image search in very large collections of images

Participants : Saloua Ouertani-Litayem, Alexis Joly, Nozha Boujemaa.

With the rapid development of information acquisition technology, we have witnessed an explosive growth in the scale of shared data collections. Then, it is now possible to tackle fundamental problems with very large datasets' context. Especially those addressing challenging tasks in machine learning for developing large scale approaches for multimedia retrieval and mining. Computer Vision is experiencing this paradigm shift, with large annotated image and video datasets becoming available. Indeed, various benchmark datasets for image classification have been released such as image-net and LabelMe. Therefore, a key challenge is taken up through out the Phd aiming to build efficient methods for training and matching efficiently very large collections of images.

We proposed several SVM-based strategies to build new supervised hash function families from large annotated collections of features. We indeed investigated with an approach consisting in benefiting from different embedding approaches in order to build compact codes indexed with efficient similarity search structures. Therefore, we have extended a kernelized hashing method [14] with multi-class SVM to solve a K-class classification problem by choosing the maximum applied to the outputs of K SVMs. We indeed proposed hashing methods based on the multi-class SVM classification strategies: One vs One (OVO) And One vs All (OVA). An important task during this process was to experimentally evaluate the quality lost induced by such representations with respect to the efficiency gains. We then compared multi-class SVM strategies with different underlying kernels.

Inspired by state of the art hashing in kernel space methods we investigated an approach consisting of benefiting from both semantic hashing like techniques and kernel embedding approach in order to build compact category aware codes indexed with efficient similarity. Experiments, are performed on image-net ILSVRC 2010 dataset [23] . Results will be submitted for publication.