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

Abstract

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.

Unsupervised similarity learning through Cartesian product of ranking references

“…The small number of images per class (only 4) makes this dataset a very challenging one for unsupervised learning algorithms. Despite this fact, the CPRR achieved high gains ranging from +5.08% to +17.72% and superior to a recent baseline [9] . Fig.…”

“…In this work, two different depths are considered: L , which defines a broader neighborhood used by the rank normalization step; and k , which defines a local neighborhood used by Cartesian product operations. Once both k and L values are much smaller than n , a sparse matrix structure [9] can be used for storage of similarity scores.…”

“…The relative effectiveness gains, com puted based on serial execution, achieve very high values up to +32.57%. The results of recent unsupervised learning approaches [9,51] are reported as baselines. We can observe that the proposed CPRR algorithm yields the best scores for most of descriptors.…”

“…As a promising alternative, rank-based approaches have attracted a lot of attention due to their capacity to take into account both effectiveness and efficiency aspects. The analysis of rankings can provide a rich and reliable source of information for contextbased measures, as demonstrated in several recent studies [7][8][9] . Different models can be exploited to analyze the rank information, such as similarity of ranked lists [7] or sets [8] , rank-based recommendations [9] , and rank consistency verifications [10] .…”

“…The analysis of rankings can provide a rich and reliable source of information for contextbased measures, as demonstrated in several recent studies [7][8][9] . Different models can be exploited to analyze the rank information, such as similarity of ranked lists [7] or sets [8] , rank-based recommendations [9] , and rank consistency verifications [10] . In general, rank-based approaches are computationally efficient, since the required computational efforts can be substantially reduced by considering only the most relevant information, which is located at top positions of the ranked lists.…”