Three things everyone should know to improve object retrieval

By Relja Arandjelovi'c, Andrew Zisserman. 2012

1,533 citations recorded by the source corpus.

Subfields: Aerospace Engineering, Computer Vision and Pattern Recognition

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

Citations in the corpus

ImageNet Large Scale Visual Recognition Challenge (2014, methodology): “…Thomas Mensink†‡, Jakob Verbeek†, Florent Perronnin‡, Gabriela Csurka‡ (Mensink et al., 2012) VGG 27.0 50.0 University of Oxford Karen Simonyan, Yusuf Aytar, Andrea Vedaldi, Andrew Zisserman (Arandjelovic and Zisserman, 2012; Sanchez et al., 2012) SuperVision 16.4 34.2 University of Toronto…”
CNN Features Off-the-Shelf: An Astounding Baseline for Recognition (2014, methodology): “mphasized that the ﬁeld of retrieval cares about both representation evaluation and the whole pipeline. The result of the whole retrieval pipelines are superior to that of representation. For example [4] achieves mAP of 0.929 on Oxford5k and Hamming Embedding [42] decreases the memory foot-print whi”
Return of the Devil in the Details: Delving Deep into Convolutional Nets (2014, methodology): “These features are square-rooted as suggested by [24], and decorrelated and reduced in dimension from 128 D to 80 D using PCA.”
Particular object retrieval with integral max-pooling of CNN activations (2015, background): “Higher scores on Paris6k are reported by Arandjelovic & Zisserman (2012) (91.0) and by Zhong et al. (2015) (91.5).”
Modeling Spatial-Temporal Clues in a Hybrid Deep Learning Framework for Video Classification (2015, methodology): “In addition, further improvements may be achieved by using advantageous feature encoding methods like the Fisher Vectors [29] or adopting feature normalization strategies, such as RootSift [1] and Power Norm [32].”
Generic 3D Representation via Pose Estimation and Matching (2016, neutral): “Matching Baselines: We compared our results with the handcrafted features of SIFT [35], Root-SIFT [7], DAISY [55], VIP [61] (which requires the surface normals in the input for which we used the normals from the 3D models), and ASIFT [41].”
InLoc: Indoor Visual Localization with Dense Matching and View Synthesis (2019, background): “This approach scales to entire cities thanks to compact image descriptors and efficient indexing techniques [6], [7],”
Understanding the Limitations of CNN-Based Absolute Camera Pose Regression (2019, methodology): “DenseVLAD densely extracts RootSIFT [4, 40] descriptors from an image and pools them into a VLAD [27] descriptor.”
Cross-Dimensional Weighting for Aggregated Deep Convolutional Features (2015, background): “Successful extensions include soft assignment [26], spatial matching [25], query expansion [6, 5, 1], better descriptor normalization [1], feature selection [35], feature burstiness [13] and very large vocabularies [21].”
SuperGlue: Learning Feature Matching With Graph Neural Networks (2019, methodology): “As observed in Section 5.3 when evaluating on the Pho-toTourism dataset, SuperGlue consistently improves over all baselines for both SIFT and SuperPoint.”
Detecting Avocados to Zucchinis: What Have We Done, and Where Are We Going? (2013, methodology): “It uses an image classification system with dense SIFT features and color statistics [13], a Fisher vector representation [16], and a linear SVM classifier, plus additional insights from [2, 17], combined with the deformable parts-based model (DPM) [9] which has been the dominant model for generic o”
Graph-Based Discriminative Learning for Location Recognition (2013, methodology): “Arandjelovic and Zisserman propose discriminative query expansion in which a per-query-image distance metric is learned based on feedback from image retrieval [2].”
“Important stuff, everywhere!” Activity recognition with salient proto-objects as context (2014, background): “It has been pointed out that the power transform of elements in a feature vector F makes the distribution of the features uniform and this way increases the discrimintive power of F [3, 29].”
Revisiting Oxford and Paris: Large-Scale Image Retrieval Benchmarking (2018, methodology): “This way, unfair comparisons are avoided in the case of methods performing off-line preprocessing of the database [2, 14]; any preprocessing should not include any part of query images.”
Exploiting local features from deep networks for image retrieval (2015, methodology): “To increase the discriminative ability of SIFT descriptors, Root-SIFT [1] was proposed to address the burstiness problem by using the Hellinger kernel on the original SIFT descriptors.”
Exemplar SVMs as visual feature encoders (2015, background): “Many successful image feature encoders operate on ad-hoc, fixed-dimensional local descriptor vectors extracted from densely [5, 1] or sparsely [19, 23] sampled local regions of the image.”