Sparsity-Induced Similarity Measure and Its Applications

Abstract

The structures of feature vectors based semi-supervised/supervised learning has gained considerable interests in the past several years thanks to its effectiveness for better object modeling and classification. In many machine learning and computer vision tasks, a critical issue is the similarity between two feature vectors. In this paper, we present a novel technique to measure the similarities among feature vectors by decomposing each feature vector as an 1 sparse linear combination of the rest of the feature vectors. The main idea is that the coefficients in such a sparse decomposition reflect the features' neighborhood structure thus providing better similarity measures among the decomposed feature vector and the rest of the feature vectors. The proposed approach is applied to label propagation and action recognition, and is evaluated on several commonly-used data sets. The experimental results show that the proposed Sparsity Induced Similarity (SIS) measure significantly improves the performance of both label propagation and action recognition.

Elastic Net subspace clustering applied to pop/rock music structure analysis

“…Since the content of a music signal is modeled by appropriate audio feature vectors, a conventional way to reveal the desired within-section similarities is to construct an SDM containing the pairwise distances between all feature vectors and then to cluster the similar feature vectors into the same music section (Dannenberg and Goto, 2008;Paulus et al, 2010). However, similarity measures, such as the Euclidean distance, the inner product, the cosine distance, and the normalized correlation, which are often used to construct the SDM for music structure analysis, ignore the subspace structure of the music sections (Cheng et al, 2012). Such subspace structures are known to be valuable for feature vector similarity measures in many clustering and classification problems (Cheng et al, 2012;Vidal, 2011;Liu et al, 2013).…”

“…However, similarity measures, such as the Euclidean distance, the inner product, the cosine distance, and the normalized correlation, which are often used to construct the SDM for music structure analysis, ignore the subspace structure of the music sections (Cheng et al, 2012). Such subspace structures are known to be valuable for feature vector similarity measures in many clustering and classification problems (Cheng et al, 2012;Vidal, 2011;Liu et al, 2013). Moreover, the aforementioned similarity measures are extremely fragile in the presence of outliers (Vidal, 2011), hindering a reliable segmentation.…”

Efficient Nearest Neighbors via Robust Sparse Hashing

“…Building similarity metrics on sparse codes has been investigated several times in the recent past ( [32], [9]). Unfortunately, the adequacy of these metrics for retrieval problems is not thoroughly investigated.…”

Structured sparse coding method for infrared small target detection in video sequence

“…In this paper, we develop a method called structured sparse coding (SSC) for infrared small target detection in video sequence. The sparse coding methods have been investigated in extensive fields [6] [7]. The main contributions of this paper are summarized as follows: (1) A collaborative convex SSC model is proposed to address the infrared small target detection problem.…”