Abstract Dynamic texture (DT) is an extension of texture to the temporal domain. How to segment DTs is a challenging problem. In this paper, we propose significant improvements to a recently published DT segmentation method. We employ a new spatiotemporal local texture descriptor which combines local binary patterns with a differential excitation measure. We also address the important problem of threshold selection by proposing a method for determining thresholds for the segmentation method by statistical learning. An improved criterion for merging adjacent regions is also introduced. Experimental results show that our approach provides very good segmentation results compared to state-of-the-art methods. 1. Introduction Dynamic textures or temporal textures are textures with motion [7, 11, 19]. There are lots of DTs in real world, including sea-waves, smoke, foliage, fire, shower and whirlwind, etc. Potential applications of DT include remote monitoring and various type of surveillance in challenging environments, such as monitoring forest fires to prevent natural disasters, traffic monitoring, homeland security applications, and animal behavior for scientific studies [4]. Segmentation is one of the basic problems in computer vision [1, 13, 16]. Meanwhile, DT segmentation is very challenging compared with the static case because of their unknown spatiotemporal extension. Examples of recent approaches are methods based on mixtures of dynamic texture model [4], mixture of linear models [9], multi-phase level sets [10], Gauss-Markov models and level sets [11], Ising descriptors [12], optical flow [20], and local binary patterns in three orthogonal planes (LBP-TOP) [5]. Although the method of [5] achieves good DT segmentation result by generalizing the frequently cited method of [14], the performances of both of these approaches depend on the threshold values needed for the splitting and merging processes. Especially the threshold for the merging process varies with regard to different textures or DTs. In this paper, we employ a new and more discriminative local texture descriptor, named Weber local descriptor (WLD) proposed by Chen et al. [6], to replace (a) (b) (c) Fig. 1. Illustration of DT segmentation; (a) splitting, (b) merging, and (c) pixelwise classification. the contrast proposed in [5]. We also address the important problem of threshold selection by proposing a method for determining thresholds by statistical learning. In addition, we improve the criterion for merging adjacent regions in DT segmentation. Experimental results demonstrate that the improved method performs favorably compared with the state-of-the-art methods. In Fig. 1, we show an example of DT segmentation using the proposed method. The contributions of this paper include:  We use the differential excitation proposed in [6] to replace the contrast used in [5] and generalize the differential excitation of a single spatial texture to a spatiotemporal mode. In [6], Chen et al. proposed a local descriptor WLD. We just use one of its components, i.e., differential excitation. However, we still call it as WLD for consistency, and its spatiotemporal mode as WLD TOP , i.e., WLD in three orthogonal planes. By combining LBP-TOP and WLD TOP we have the generalized feature (LBP/WLD) TOP .  We propose a learning based method to determine thresholds statistically instead of the experimental thresholds in [5].  We modify the merging criterion of [5]. The modified criterion requires that the two adjacent blocks should satisfy the following two conditions: 1) the similarities between these two blocks are larger than the learned thresholds; 2) the merger importance (MI) is minimal in those block pairs which satisfy the first condition. The rest of this paper is organized as follows: In Section 2, we present the feature (LBP/WLD) TOP and discuss how to use it for DT segmentation. In Section 3, we describe how to learn the thresholds. In Section 4, we show the detailed process of segmentation. In Section 5, some experimental results are presented, followed by a discussion in Section 6. An Improved Local Descriptor and Threshold Learning for Unsupervised Dynamic Texture Segmentation Jie Chen Guoying Zhao Matti Pietikäinen Machine Vision Group, Department of Electrical and Information Engineering, University of Oulu, Finland {jiechen, gyzhao, mkp}@ee.oulu.fi