Dynamic Texture Recognition Using 3D Random Features

Abstract

In this paper, we present a novel, simple but effective approach for dynamic texture recognition using 3D random features. Compared with the existing dynamic texture recognition approaches using carefully designed features for high performance, our method use only a few 3D random filters to extract spatio-temporal features from local dynamic texture blocks, which are further encoded into a low-dimensional feature vector. To explore the representative power of the 3D random features, we use two different encoding schemes, the learning-based Fisher vector encoding and the learning-free binary encoding. The proposed method is tested on the UCLA and DynTex databases with various evaluation protocols. Experimental results demonstrate the high performance of our method for dynamic texture recognition.

Dynamic Texture Classification Using Directional Binarized Random Features

“…Fisher vector [86] encoding has also been used to build dictionaries from DTs. Zhao et al [23] first extracted 3D random features from sampled 3D blocks and then trained a Gaussian mixture model (GMM), after which Fisher vector encoding was applied to generate DT feature vectors (denoted as 3DRF). Later, Xiong et al [24] extended 3DRF by replacing the random filters with those learned via independent component analysis.…”

“…After obtaining a bunch of vectors of L random features from a DT, the task is to aggregate those local features into a global feature vector that could be used for DT classification. Several commonly used techniques are dictionary learning [26], [92], [93], k-means clustering [96], Fisher vector encoding [23], and histogramming on local binary codes [22], [25]. The former three techniques require a training process to generate a dictionary or a set of cluster centers (used as a dictionary), which is not in line with our goal.…”

“…• Gamma: This subset contains 275 videos, which also belong to ten classes of flowers (29), sea (38), naked trees (25), foliage (35), escalator (7), calm water (30), flags (31), grass (23), traffic (9), and fountains (37).…”

“…They later proposed to view a DT along three axes (i.e., the horizontal, vertical, and temporal axes) and treated it as three image sequences, from which LBP features were extracted and concatenated as the final feature vector (denoted as LBP-TOP) [4]. VLBP and LBP-TOP turned out to be a great success for DT classification and started a thread of research that applies binary encoding in a local neighborhood for DT description [5], [14], [15], [16], [17], [18], [19], [20], [21], [22], [23], [24].…”

Learning graph representation with Randomized Neural Network for dynamic texture classification

“…However, they have known limitations that can make them unfeasible for many real-world problems, such as their difficulty to be implemented on embedded platforms, and the need for a considerable number of training samples, that can be impossible to get, particularly when online adaptation/learning is needed. Zhao et al [34] explore two different approaches to learn 3D random features: learning-based Fisher vector and the learning-free binary encoding. In [35] is proposed a DT descriptor, which employs Randomized Neural Networks (RNNs) to learn the local features from three orthogonal planes.…”