Image Re-ranking Acceleration on GPUs

Concurrency and Computation

Torres

et al. 2016

Self Cite

Re-ranking algorithms have been proposed to improve the effectiveness of content-based image retrieval systems by exploiting contextual information encoded in distance measures and ranked lists. In this paper, we show how we improved the efficiency of one of these algorithms, called Contextual Spaces Re-Ranking (CSRR). One of our approaches consists in parallelizing the algorithm with OpenCL to use the central and graphics processing units of an accelerated processing unit. The other is to modify the algorithm to a version that, when compared with the original CSRR, not only reduces the total running time of our implementations by a median of 1:6 but also increases the accuracy score in most of our test cases. Combining both parallelization and algorithm modification results in a median speedup of 5:4 from the original serial CSRR to the parallelized modified version. Different implementations for CSRR's Re-sort Ranked Lists step were explored as well, providing insights into graphics processing unit sorting, the performance impact of image descriptors, and the trade-offs between effectiveness and efficiency. vectors, which are later used for similarity assessment among images. A CBIR system ranks the collection images by decreasing order of similarity, and because users consider mostly top-ranked images, it is imperative that the rank be as accurate as possible.In recent years, several successful attempts to increase the effectiveness (quality of results) of CBIR systems have been performed [1][2][3][4][5][6]. In particular, re-ranking methods have been used to improve the effectiveness of CBIR systems by exploiting contextual information encoded in similarity scores and ranked lists. These methods are, on the other hand, very costly as they are based on comparing collection images multiple times. In a real-world scenario, CBIR systems require both good effectiveness and efficiency (response time), so re-ranking methods must be improved.Central processing units (CPUs) no longer have just one core, and graphics processing units (GPUs) are now being used as general purpose processors due to having evolved into massive parallel architectures capable of executing hundreds of operations per cycle [7]. These devices have been successfully used to accelerate re-ranking [7, 8] and retrieval [9] systems, obtaining good speedups.Therefore, alternatives that increase performance with parallelization seem to be a possible fit for the Contextual Spaces Re-Ranking (CSRR) algorithm [2], which we discuss in this paper. Another possible approach consists in analyzing the compromises between accuracy and performance that come out of modifying existing algorithms, because this can lead to eliminating demanding work.Our solution first exploits the use of parallelization to speed up the more costly steps of the CSRR algorithm, obtaining speedups of up to 3:3 for the Compute Distances step and 5:1 for the Re-sort Ranked Lists step on an accelerated processing unit (APU). We then propose a modification to this algorithm, which by itse...

Section: Introductionmentioning

confidence: 99%

mentioning

confidence: 99%

Contextual Spaces Re‐Ranking: accelerating the Re‐sort Ranked Lists step on heterogeneous systems

Pisani

Concurrency and Computation

Torres

et al. 2016

Self Cite

“…wise Recommendation [28] and the RL-Sim algorithms [29,31] are considered as baselines. The position of algorithms in the graph is given by the MAP score and the run time.…”

Section: Effectiveness Evaluationmentioning

confidence: 99%

“…The reported results do not consider the OpenCL build and environment time, since the build can be executed once off-line and the environment time is constant independently of dataset sizes. We also present a comparison with other two unsupervised learning algorithms: the Pairwise Recommendation [28] and the RL-Sim algorithms [29,31]. Table 5 presents the average run time and confidence intervals for the RL-Recommendation algorithm considering different criteria.…”

Section: Efficiency Evaluationmentioning

confidence: 99%

See 1 more Smart Citation

Effective, Efficient, and Scalable Unsupervised Distance Learning in Image Retrieval Tasks

Valem

Proceedings of the 5th ACM on International Conference on Multimedia Retrieval

Torres

et al. 2015

Self Cite

Various unsupervised learning methods have been proposed with significant improvements in the effectiveness of image search systems. However, despite the relevant effectiveness gains, these approaches commonly require high computation efforts, not addressing properly efficiency and scalability requirements. In this paper, we present a novel unsupervised learning approach for improving the effectiveness of image retrieval tasks. The proposed method is also scalable and efficient as it exploits parallel and heterogeneous computing on CPU and GPU devices. Extensive experiments were conducted considering five different public image collections and several descriptors. This rigorous experimental protocol evaluates the effectiveness, efficiency, and scalability of the proposed approach, and compares it with previous methods. Experimental results demonstrate that high effectiveness gains (up to +29%) can be obtained requiring small run times.

A unified model for accelerating unsupervised iterative re‐ranking algorithms

Pisani

Valem

Concurrency and Computation

et al. 2020

Self Cite

Re-ranking algorithms are used to improve the effectiveness of multimedia retrieval systems. However, they are usually very computationally costly, and therefore demand the specification and implementation of efficient and effective big multimedia analysis approaches.Recently proposed unsupervised iterative re-ranking methods present good accuracy and significant potential for parallelization, leading us to explore efficiency vs. effectiveness trade-offs. In this paper, we introduce a class of unsupervised iterative re-ranking algorithms and present a model that can be used to guide their implementation for parallel architectures. We also analyze the impact of the parallelization on the performance of three algorithms that belong to the proposed class: Contextual Spaces, RL-Sim, and Contextual Re-Ranking. The experiments show speedups that reach up to 6.0× for Contextual Spaces Re-Ranking, 16.1× for RL-Sim Re-Ranking, and 3.3× for Contextual Re-Ranking. These results demonstrate that the proposed parallel programming model can be successfully applied to improve the scalability of multimedia retrieval systems.