All About VLAD

By Relja Arandjelovi'c, Andrew Zisserman. 2013 IEEE Conference on Computer Vision and Pattern Recognition, 2013

834 citations recorded by the source corpus.

The objective of this paper is large scale object instance retrieval, given a query image. A starting point of such systems is feature detection and description, for example using SIFT. The focus of this paper, however, is towards very large scale retrieval where, due to storage requirements, very compact image descriptors are required and no information about the original SIFT descriptors can be accessed directly at run time. We start from VLAD, the state-of-the art compact descriptor introduced by Jegou et al. for this purpose, and make three novel contributions: first, we show that a simple change to the normalization method significantly improves retrieval performance, second, we show that vocabulary adaptation can substantially alleviate problems caused when images are added to the dataset after initial vocabulary learning. These two methods set a new state-of-the-art over all benchmarks investigated here for both mid-dimensional (20k-D to 30k-D) and small (128-D) descriptors. Our third contribution is a multiple spatial VLAD representation, MultiVLAD, that allows the retrieval and localization of objects that only extend over a small part of an image (again without requiring use of the original image SIFT descriptors).

Subfields: Aerospace Engineering, Computer Vision and Pattern Recognition

View Paper: https://www.semanticscholar.org/paper/25d0fa49ca846370ff4796a6ac6688a42cf50f77

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

24/7 Place Recognition by View Synthesis (2018, positive): “nine times more images than in the original street-view— we use the compact VLAD encoding [25], [26] of local image descriptors, which is amenable to efficient compression, storage and indexing.”
CNN Features Off-the-Shelf: An Astounding Baseline for Recognition (2014, methodology): “8[20] VLAD[4] 32k 55.”
Geometry-Aware Learning of Maps for Camera Localization (2017, methodology): “, bag-of-words (BoW) features [12], VLAD [4], Fisher vectors [28], and recent DNN-based features [3]) are used to build maps/vocabularies for image retrieval and pose estimation.”
FSA-Net: Learning Fine-Grained Structure Aggregation for Head Pose Estimation From a Single Image (2019, background): “Facial modeling and analysis have long been an active research topic in computer vision [2, 3, 4, 5, 6, 7, 21, 22, 24, 25].”
Efficient Global 2D-3D Matching for Camera Localization in a Large-Scale 3D Map (2017, background): “, L1 norm distances between Bag-of-Words vectors [19], L2 norm distances between compact representations/vectors [6, 7]), and the position of the best database image is deemed as that of the query image or the top N images are fused to get the position of the query image [46, 47, 50].”
Multi-view Super Vector for Action Recognition (2014, methodology): “We then apply intra-normalization for each component as in [2].”
Fine-Tuning CNN Image Retrieval with No Human Annotation (2017, background): “Previous compact descriptors are typically constructed by an aggregation of local features, where representatives are Fisher vectors [28], VLAD [29] and alternatives [30], [31], [32].”
Deep Face Recognition (2015, methodology): “This is done by computing the VLAD descriptor [2, 9] for each image, clustering such descriptors within the 1,000 images for each identity using a very tight threshold, and retaining a single element per cluster.”
CNN Image Retrieval Learns from BoW: Unsupervised Fine-Tuning with Hard Examples (2016, methodology): “Representative approaches are Fisher vectors [9], VLAD [10] and alternatives [11,12,13].”
Predicting Good Features for Image Geo-Localization Using Per-Bundle VLAD (2015, methodology): “We tested multiple normalization schemes [1, 19], but the combination of residual- and L2- normalization performed the best in our data.”
From single image query to detailed 3D reconstruction (2015, methodology): “Most of the approaches pose the problem as a nearest neighbor search in a descriptor space, such as bag-of-words [28, 21, 25, 11], VLAD [12, 2], Fischer vectors [24], or exhaustive matching [34, 27].”
A discriminative CNN video representation for event detection (2014, methodology): “Except for the power and l2 normalization, we apply intranormalization [4] to VLAD.”
Grassmann Pooling as Compact Homogeneous Bilinear Pooling for Fine-Grained Visual Classification (2018, methodology): “To further suppress bursts, another kind of normalization termed intra-normalization was proposed in [27], where the sum of residuals is L2 normalized within each VLAD block.”
Learned Contextual Feature Reweighting for Image Geo-Localization (2017, background): “Following [2], f is intra-normalized [3], then L2 normalized.”
Encoding High Dimensional Local Features by Sparse Coding Based Fisher Vectors (2014, methodology): “In this paper, we apply intra-normalization [12] to normalize the pooled Fisher vectors.”
To Aggregate or Not to aggregate: Selective Match Kernels for Image Search (2013, neutral): “These differences are summarized in Table 1.”