Efficient nearest neighbors search for large scale
- 1. Efficient Nearest Neighbors Search for Large-Scale Landmark Recognition Federico Magliani, Tomaso Fontanini, and Andrea Prati IMP Lab - University of Parma 22/10/2018 IMP Lab - University of Parma 1
- 2. Agenda • Motivations • Related works • Proposed approach (Bag of Indexes) • Experimental results • Conclusions 22/10/2018 IMP Lab - University of Parma 2
- 3. Motivations • Approximate Nearest Neighbor (ANN) search problem • find relevant results among an huge quantity of data • trade-off between computational time and memory occupancy • applied on image, text and information retrieval 22/10/2018 IMP Lab - University of Parma 3
- 4. Agenda • Motivations • Related works • Proposed approach (Bag of Indexes) • Experimental results • Conclusions 22/10/2018 IMP Lab - University of Parma 4
- 5. Related works • Permutation Pivots allows to represent the image descriptors through permutation of a set of randomly selected reference objects; • Locality Sensitive Hashing (LSH) projects points that are close to each other into the same bucket with high probability; • Product Quantization (PQ) decomposes the space into a Cartesian product of low dimensional subspaces and quantizes each subspace separately; • FLANN: an open source library for ANN and one of the most popular for nearest neighbor matching. 22/10/2018 IMP Lab - University of Parma 5
- 6. Agenda • Motivations • Related works • Proposed approach (Bag of Indexes) • Experimental results • Conclusions 22/10/2018 IMP Lab - University of Parma 6
- 7. Proposed approach: Bag of Indexes (BoI) It’s a multi-index hashing algorithm for ANN search problem. • The Db data are projected through LSH function and the index of the signature is saved in hash tables; • For each query, the following process is repeated for every projection: 1. Project the descriptor. 2. The indexes found in the bucket closest to the query will be added to a ranking list (BoI) with a weight proportional by the Hamming distance between the query bucket and the analysed bucket. 3. At the end the topN elements are re-ranked according to the Euclidean distance. 22/10/2018 IMP Lab - University of Parma 7
- 8. Proposed approach: Bag of Indexes (BoI) 0 1 2 3 1 2 3 4 5 6 7 Weight Image Index Hash Table 1 Hash Table 2 Hash Table 3 22/10/2018 IMP Lab - University of Parma 8 Hash Table 1 … … {4,6} 5 {2,3} … … Hash Table 2 … 7 {5,3} 1 … … … Hash Table 3 … … … 5 3 {1,4} … Index of query image for each Hash Table L = 3
- 9. Proposed approach: Bag of Indexes (BoI) • Weighing strategy (multi-probe approach): 𝑤 𝑖, 𝑞, 𝑙 = ቐ 1 2 𝐻(𝑖,𝑞) , 𝑖𝑓 𝐻 𝑖, 𝑞 ≤ 𝑙 0, 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒 where i is a generic bucket, q is the query bucket and H(i,q) is the Hamming distance between i and q. • Adaptive version: after a predefined number of hash table, the gap is reduced in order to reduce the computational time. 22/10/2018 IMP Lab - University of Parma 9
- 10. Linear vs Sublinear reduction • linear: the number of neighboring buckets γ is reduced by 2 every 40 hash tables: 𝛾𝑖 = ቊ 𝛾𝑖−1 − 2, 𝑖𝑓 𝑖 = {Δ1, … , 𝑘𝑖Δ1} 𝛾𝑖−1, 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒 with i = {1, . . . , L}, ∆1 = 40 and k1 : k1 ∆1 ≤ L • sublinear: the number of neighboring buckets γ is reduced by 2 every 25 hash tables, but only after the first half of hash tables: 𝛾𝑖 = 𝛾𝑖−1, 𝑖𝑓 𝑖 ≤ 𝐿 2 𝛾𝑖−1 − 2, 𝑖𝑓 𝑖 = 𝐿 2 , 𝐿 2 + Δ2, … , 𝐿 2 + 𝑘2Δ2 𝛾𝑖−1, 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒 with i = {1, . . . , L}, ∆2 = 25 and k2 : L/2 + k2∆2 ≤ L 22/10/2018 IMP Lab - University of Parma 10
- 11. BoI - Parameters config Symbol Name Value δ hash dimension 2 8 = 256 L hash tables 100 𝜸 𝟎 initial gap 68 l neighbors used 3-neighbors - reduction sublinear ε re-ranking top 250 elements 22/10/2018 IMP Lab - University of Parma 11
- 12. Agenda • Motivations • Related works • Proposed approach (Bag of Indexes) • Experimental results • Conclusions 22/10/2018 IMP Lab - University of Parma 12
- 13. Datasets • Holidays+Flickr1M (1M distractor images + 1491 images: 500 classes, 500 query.) • Oxford105k (100k distractor images + 5062 images: 11 classes, 55 queries); • Paris106k (100k distractor images + 6412 images: 11 classes, 55 queries); • SIFT1M (1M 128D SIFT descriptors, 10k query images, only the top 100 images in the final ranking for each query are evaluated) • GIST1M (1M 960D GIST descriptors, 1k query images, only the top 100 images in the final ranking for each query are checked) 22/10/2018 IMP Lab - University of Parma 13
- 14. Evaluation Metrics • Different evaluation metrics are used to compare with the state-of- the-art approaches: • Recall in R = 1, 10, 100 → it is the average rate of queries for which the 1-nearest neighbor is ranked in the top R positions. • mAP (mean Average Precision) → mean of Average Precision scores (correct results) for each query, based on the position in the ranking. 22/10/2018 IMP Lab - University of Parma 14
- 15. Results on Holidays+Flickr1M Method ε mAP Avg retrieval time (msec) LSH 250 86.03 % 3103 Multi-probe LSH 250 86.10 % 16706 Permutations 250 82.70 % 2844 LOPQ 250 36.37 % 4 FLANN 250 83.97 % 995 BoI LSH 250 78.10 % 5 BoI multi-probe LSH 250 85.16 % 12 BoI adaptive multi-probe LSH 250 85.35 % 8 22/10/2018 IMP Lab - University of Parma 15
- 16. Results on Holidays+Flickr1M Method ε mAP Avg retrieval time (msec) Permutations 10k 85.51 % 15640 LOPQ 10k 67.22 % 72 FLANN 10k 85.66 % 1004 BoI adaptive multi-probe LSH 10k 86.09 % 16 22/10/2018 IMP Lab - University of Parma 16
- 17. Results on Oxford105k and Paris106k Method ε Oxford105k Paris106k mAP Avg ret. Time (msec) mAP Avg ret. Time (msec) LSH 2500 80.83% 610 86.50% 607 Permutations 2500 81.89% 240 88.14% 140 LOPQ 2500 71.70% 346 87.47% 295 FLANN 2500 70.33% 2118 68.93% 2132 Boi adaptive multi-probe LSH 2500 81.44% 12 87.90% 13 Permutations 10k 82.82% 250 89.04% 164 LOPQ 10k 69.94% 1153 88.00% 841 FLANN 10k 69.37% 2135 70.73% 2156 Boi adaptive multi-probe LSH 10k 84.38% 25 92.31% 26 22/10/2018 IMP Lab - University of Parma 17
- 18. Results on Sift1M Method ε R=1 R=10 R=100 Avg retrieval time (msec) Permutations 500 94.32 % 94.98% 94.98 % 16999 LOPQ 500 19.93 % 44.80 % 52.92 % 3 FLANN 500 54.47 % 54.83 % 54.83% 16 BoI adaptive multi-probe LSH 500 93.72 % 94.34 % 94.34 % 22 LOPQ 10k 36.34 % 80.11 % 96.18 % 104 FLANN 10k 95.06 % 95.86 % 95.86 % 31 BoI adaptive multi-probe LSH 10k 99.17 % 99.85 % 99.85 % 30 22/10/2018 IMP Lab - University of Parma 18
- 19. Results on Gist1M Method ε R=1 R=10 R=100 Avg retrieval time (msec) Permutations 500 54.80 % 55.30% 55.30 % 17909 FLANN 500 28.30 % 28.60 % 28.60% 1262 BoI adaptive multi-probe LSH 500 57.70 % 58.20 % 58.20 % 69 LOPQ 10k 75.90 % 76.50 % 76.50 % 1352 BoI adaptive multi-probe LSH 10k 92.40 % 93.40 % 93.40 % 108 22/10/2018 IMP Lab - University of Parma 19
- 20. Agenda • Motivations • Related works • Proposed approach (Bag of Indexes) • Experimental results • Conclusions 22/10/2018 IMP Lab - University of Parma 20
- 21. Conclusions • The proposed Bag of Indexes (BoI) adaptive multi-probe LSH is a simple technique implemented for the efficient resolution of the ANN search problem. • BoI allows to work in combination of different hashing/projection functions. • Experiments are performed on five public datasets, namely Holidays+Flickr1M, Oxford105k, Paris106k, SIFT1M and GIST1M, and demonstrate superior recognition accuracy w.r.t. the state of the art. 22/10/2018 IMP Lab - University of Parma 21
- 22. Thanks for your attention! • Questions? • Contacts: tomaso.fontanini@studenti.unipr.it • Website: implab.ce.unipr.it/?page_id=122 • GitHub: github.com/fmaglia/BoI 22/10/2018 IMP Lab - University of Parma 22