Sampling Strategies for Bag-of-Features Image Classification

By Eric Nowak, Frédéric Jurie, Bill Triggs. 2006

1,155 citations recorded by the source corpus.

Subfields: 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 Computer Vision and Pattern Recognition

Citations in the corpus

Deep learning in neural networks: An overview (2014, methodology): “, 2009) using a bag of features approach (Nowak et al., 2006) to extract regions of interest.”
Visualizing Deep Convolutional Neural Networks Using Natural Pre-images (2015, methodology): “Here we consider two densely-sampled versions (Nowak et al. 2006), namely DSIFT (Dense SIFT) and HOG (Dalal and Triggs 2005).”
An Analysis of Single-Layer Networks in Unsupervised Feature Learning (2011, methodology): “For instance, it is widely held that highly over-complete feature representations tend to give better performance than smaller-sized representations [32], and similarly with small strides between features [21].”
Fisher Vector Faces in the Wild (2013, background): “Dense feature extraction is an essential component of many state-of-the-art image classification methods [18, 23, 27].”
Object categorization using co-occurrence, location and appearance (2008, neutral): “BoF is a widely used discriminative model for recognition [7, 16].”
Visual Word Ambiguity (2010, methodology): “Typically, a vocabulary is constructed by applying k-means clustering on image features [4], [14]– [16], [20], [23]–[26], [30], [31].”
A Bag-of-Features Framework to Classify Time Series (2013, methodology): “BoF representations are popular, mostly in computer vision as content-based image retrieval [21], [22], [23], natural scene classification [24], and object detection and recognition [25], [26], [27], [28], [29] because of their simplicity and good performance [30].”
Multiple instance classification: Review, taxonomy and comparative study (2013, methodology): “In this work we use KM, as it is the most widely used [5, 30].”
Feature sampling and partitioning for visual vocabulary generation on large action classification datasets (2014, methodology): “These include the patch sampling strategy, for which uniform random sampling of the image space was shown to outperform interest point-based samplers such as Harris-Laplace and Laplacian of Gaussian [7].”
Multi-Region Probabilistic Histograms for Robust and Scalable Identity Inference (2009, methodology): “The use of probabilistic histograms in MRH differs to the histograms used in [4] (for image retrieval/categorisation purposes), where a Vector Quantiser (VQ) based strategy is typically used.”
Pyramid of Spatial Relatons for Scene-Level Land Use Classification (2015, methodology): “In this paper, we apply SIFT feature on gray images with densely sampled interest points, which is the most popular approach in both computer vision and remote sensing communities [28], [30], [42]–[44].”
Randomized Clustering Forests for Image Classification (2008, background): “This means that any “weak” information also improves the performance, which confirms the conclusion in [42] about sampling.”
Dense interest points (2010, background): “[13, 4 ]), but this does not seem to have any advantage over regular grids, and equally ignores the actual image content during the feature”
Data-driven grasping with partial sensor data (2009, methodology): “Unlike [27] we use a binary bag of features rather than a histogram, using the simple absence/presence rule that was shown to perform well in [26].”