PROJECT TITLE :

High-Order Local Pooling And Encoding Gaussians Over A Dictionary Of Gaussians. - 2017

ABSTRACT:

Native pooling (LP) in configuration (feature) space proposed by Boureau et al. explicitly restricts similar features to be aggregated, which will preserve as abundant discriminative information as possible. At the time it appeared, this methodology combined with sparse coding achieved competitive classification results with solely a small dictionary. However, its performance lags so much behind the state-of-the-art results as only the zero-order info is exploited. Inspired by the success of high-order statistical information in existing advanced feature coding or pooling strategies, we create an attempt to deal with the limitation of LP. To this finish, we tend to gift a completely unique method known as high-order LP (HO-LP) to leverage the information beyond the zero-order one. Our plan is intuitively easy: we compute the primary- and second-order statistics per configuration bin and model them as a Gaussian. Accordingly, we have a tendency to use a collection of Gaussians as visual words to represent the universal likelihood distribution of options from all classes. Our drawback is naturally formulated as encoding Gaussians over a dictionary of Gaussians as visual words. This problem, but, is difficult since the area of Gaussians isn't a Euclidean area but forms a Riemannian manifold. We have a tendency to address this challenge by mapping Gaussians into the Euclidean house, that allows us to perform coding with common Euclidean operations instead of advanced and often expensive Riemannian operations. Our HO-LP preserves the benefits of the initial LP: pooling solely similar features and employing a tiny dictionary. Meanwhile, it achieves very promising performance on commonplace benchmarks, with either conventional, hand-designed features or deep learning-based mostly features.