Efficient representation of local geometry for large scale object retrieval

By Michal Perdoch, Ondřej Chum, Jiri Matas. 2009

310 citations recorded by the source corpus.

Subfields: Aerospace Engineering, Computer Vision and Pattern Recognition, Information Retrieval & Search

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Aerospace Engineering

In Computer Vision and Pattern Recognition

In Information Retrieval & Search

Citations in the corpus

Product Quantization for Nearest Neighbor Search (2011, methodology): “Note that some geometrical information can also be inserted in this entry, as proposed in [18] and [25].”
End-to-End Learning of Deep Visual Representations for Image Retrieval (2016, background): “An advantage of these techniques is that spatial verification can be employed to rerank a shortlist of results (Philbin et al. 2007; Perdoch et al. 2009), yielding a large improvement despite a significant cost.”
Image Retrieval for Image-Based Localization Revisited (2012, background): “In contrast, the inverted file indices and feature geometry needed for image retrieval can be stored very efficiently [15, 24].”
SIFT Meets CNN: A Decade Survey of Instance Retrieval (2016, methodology): “Most SIFT-based methods employ the original images, except [32], [71] which manually rotate the images into upright orientations.”
Geometric min-Hashing: Finding a (thick) needle in a haystack (2009, methodology): “In particular, we use hessian affine features and the SIFT descriptor [16] with gravity vector – details can be found in [20].”
VocMatch: Efficient Multiview Correspondence for Structure from Motion (2014, methodology): “The vocabulary has 2 layers (4,096×4,096 words) and was trained on a database of 11 billion SIFT descriptors computed at Hessian-Affine interest points [16].”
Searching with quantization: approximate nearest neighbor search using short codes and distance estimators (2009, background): “Note that some geometrical information can also be inserted in this entry, as in [16] and [26].”
Location Recognition Using Prioritized Feature Matching (2010, background): “Much of this recent work has posed the problem as that of image retrieval [1–4]: given a query image to be recognized, find a set of similar images from a database using image descriptors or visual features (possibly with a geometric verification step), often building on bag-of-words techniques [5–9”
Image retrieval with geometry-preserving visual phrases (2011, methodology): “Spatial verification methods [15] and query expansion [5] have been proposed for re-ranking at the post-processing step, and many methods [9, 22, 8, 14] have been introduced to decrease the memory usage for the inverted files.”
Retrieving landmark and non-landmark images from community photo collections (2010, methodology): “Rigid geometry is 4 http://en.wikipedia.org/wiki/Great-circle distance typically veriﬁed by means of RANSAC [7] or simpler forms of spatial matching like [26], [25].”
Fine-Grained Categorization by Alignments (2013, methodology): “Additionally, we follow the gravity vector assumption [22] and adopt the highest end point of the principal axis as its origin.”
From single image query to detailed 3D reconstruction (2015, methodology): “Hessian affine features [17] are detected (1900 features per image on average) and described by the rotation-variant [23] SIFT descriptor [16].”
To Aggregate or Not to aggregate: Selective Match Kernels for Image Search (2013, unknown): “[18], which includes the gravity vector assumption and improves retrieval performance.”
Revisiting Oxford and Paris: Large-Scale Image Retrieval Benchmarking (2018, methodology): “We did not consider geometry quantization [31].”
Local Alignments for Fine-Grained Categorization (2014, methodology): “tor assumption (Perdóch et al., 2009) of the green arrows, we obtain the dominant pose orientation, see red arrows.”
Handling Urban Location Recognition as a 2D Homothetic Problem (2010, background): “It can therefore be diﬃcult to ﬁnd strong geometrical constraints for ﬁltering the initial visual-word based results, although recent approaches look at (locally) consistent orientations and feature shapes [16,17,18] and exploit that pictures are usually not taken upside down.”
Learning a Fine Vocabulary (2010, background): “While the reported bits per feature required in the search index ranges from 11 bits [8] to 18 bits [11], hamming embedding adds another 64 bits.”