A one-pass resource-allocating codebook for patch-based visual object recognition

Abstract

Frequencies of occurrence of low-level image features is the representation of choice in the design of state-of-theart visual object recognition systems. A crucial step in this process is the construction of a codebook of visual features, which is usually done by cluster analysis of a large number of low-level image features detected as interest points. However, clustering is a process that retains regions of high density in a distribution and it follows that the resulting codebook need not have discriminant properties. Here we extend our recent work on constructing a one-pass discriminant codebook design procedure inspired by the resource allocating network model from the artificial neural networks literature. Unlike clustering, this approach retains data spread out more widely in the input space, thereby including rare low-level features in the codebook. It simultaneously achieves increased discrimination and a drastic reduction in the computational needs. We illustrate some properties of our method and compare it to a closely related approach.

A Review of Codebook Models in Patch-Based Visual Object Recognition

“…The overall error rate of the classification is 15% using SVMs. Our resourceallocating codebook (RAC) approach in [50] when applied on the Xerox7 image dataset performs slightly better than the authors' method but was achieved in a tiny fraction of computation time. -Jurie and Triggs [19] proposed a mean-shift based clustering approach to construct codebooks in an undersampling framework.…”

“…The bag-of-words (BOW) approach was originally used in text mining [17] and is now widely used in image scene classification [13,48], retrieval of objects from a movie [52], and object classification [9,19,43,50,59,64] tasks in computer vision. The bag-of-words in computer vision is normally referred as 'bag-of-features' or 'bagof-keypoints'.…”

“…Furthermore, clustering millions of data vectors of higher dimensions into thousands of cluster centres using the K-means or GMMs techniques is not straightforward to apply and need for novel approaches arises. Therefore, our own work on a novel one-pass algorithm [50] constructs a visual codebook resulting in a codebook that retains rare and novel features while drastically reducing computational costs.…”

“…A technique that has many parameters can overfit data and generalise poorly [46]. In contrast, the RAC approach pursued in [50] has a single threshold that takes only one-pass through the entire data, making it computationally efficient. -Nister and Stewenius [42] proposed a hierarchical K-means clustering that constructs a vocabulary tree in an offline training stage for image retrieval from a large database.…”

“…The authors empirically show that the K-median codebook is a compromise between the HIK and K-means codebooks. -In our recent work [50], we demonstrated a resource-allocating codebook (RAC) method to constructing a discriminant visual codebook that takes only one-pass through the entire data, inspired by the resource allocation network (RAN) family of algorithms [45]. The RAC starts by arbitrarily assigning the first data item as an entry in the codebook.…”