Bundling Features for Large Scale Partial-Duplicate Web Image Search Zhong Wu ∗, Qifa Ke, Michael Isard, and Jian Sun CVPR 2009
Outline Introduction Bundled features Image Retrieval using bundled feature Experiments and results Conclusion
Outline Introduction Bundled features Image Retrieval using bundled feature Experiments and results Conclusion
Target Given a query image, is to locate its near- and partial-duplicate images in a large corpus of web images.
Unlike object-based image retrieval
State-of-the-art Visual word(quantization) & scalable textual index retrieval schemes Post-processing – Geometric verification Bundled feature – Weak geometric verification Bundled feature = SIFT + SMER
Outline Introduction Bundled features Image Retrieval using bundled feature Experiments and results Conclusion
MSER Maximally Stable Extremal Region
MSER
Bundled features
Discriminative power Increase discriminative power – Feature region size – Feature dimensionality Drawbacks – Less repeatable – Localization accuracy – Sensitive to occlusion, photometric, geometric
Matching bundled features
Bundled features
Advantage More discriminative Allowed to have large overlap error – Partially match Robust – Occlusion – Geometric changes – …etc
Outline Introduction Bundled features Image Retrieval using bundled feature Experiments and results Conclusion
Feature quantization Hierarchical k-means – One million visual words from 50K training images
Feature quantization K-D tree – pointList = [(2,3), (5,4), (9,6), (4,7), (8,1), (7,2)]
Matching bundled features
Matching bundled features
Inverted-file index Documents – T 0 = "it is what it is" – T 1 = "what is it" – T 2 = "it is a banana" Index – "a": {2} – "banana": {2} – "is": {0, 1, 2} – "it": {0, 1, 2} – "what": {0, 1}
Indexing and retrieval Support – 512 bundled features each image – 32 visual word each bundled feature
Indexing and retrieval Voting
Indexing and retrieval tf – 100 vocabularies in a document, ‘a’ 3 times – 0.03 (3/100) idf – 1,000 documents have ‘a’, total number of documents 10,000,000 – 9.21 ( ln(10,000,000 / 1,000) ) if-idf = 0.28( 0.03 * 9.21)
Outline Introduction Bundled features Image Retrieval using bundled feature Experiments and results Conclusion
Dataset Basic dataset – One million images most frequently clicked in a popular commercial image-search engine – (50K, 200K, 500K) Ground truth – Manually labeled 780 partial-duplicate web image form 19 groups. – Evaluation dataset = basic dataset + ground truth Query – 150 images from ground truth
mAP Mean average precision EX: – two images A&B – A has 4 duplicate images – B has 5 duplicate images – Retrieval rank A: 1, 2, 4, 7 – Retrieval rank B: 1, 3, 5 – Average precision A = (1/1+2/2+3/4+4/7)/4=0.83 – Average precision B = (1/1+2/3+3/5+0+0)/3=0.45 – mAP= ( )/2=0.64
Evaluation Baseline – Bag-of-features approach with soft assignment[13] [13] J. Philbin, O. Chum, M. Isard, J. Sivic, and A. Zisserman. Lost in quantization: Improving particular object retrieval in large scale image databases. In CVPR, 2008.
Evaluation Compare(HE) – enhance the with hamming embedding [3] by adding a 24-bit hamming code to filter out target features. [3] H. Jegou, M. Douze, and C. Schmid. Hamming embedding and weak geometric consistency for large scale image search. In ECCV, 2008.
Evaluation baseline0.35 to Bundled(mem)0.40 a 14% improvement baseline0.35 to Bundled 0.49 a 40% improvement baseline0.35 to Bundled+HE0.52 a 49% improvement
Evaluation Compare(Re-ranking) – Full geometric verification, RANSAC for top 300 candidate images
Evaluation Baseline+re-rank 0.50 to Bundled+re-rank 0.62 a 24% improvement Baseline 0.35 to Bundled+re-rank 0.62 a 77% improvement
Evaluation Trade-off Run time – a single CPU on a 3.0GHz Core Duo desktop with 16G memory
Sample results AP from 0.51 to 0.74 a 45% improvement
Sample results
Outline Introduction Bundled features Image Retrieval using bundled feature Experiments and results Conclusion
Bundled features for large scale partial- duplicate web image search. Bundled features property – More discriminative than individual SIFT features. – Simple and robust geometric constraints – Partially match two groups of SIFT features Advantage – Robustness to occlusion, photometric and geometric changes