Texture as a pixel feature for video object segmentation
چکیده
As texture represents one of the key perceptual attributes of any object, integrating textural information into existing video object segmentation frameworks affords the potential to achieve semantically improved performance. While object segmentation is fundamentally pixel-based classification, texture is normally defined for the entire image, which raises the question of how best to directly specify and characterise texture as a pixel feature. This letter introduces a generic strategy for representing textural information so it can be seamlessly incorporated as a pixel feature into any video object segmentation paradigm. Both numerical and perceptual results upon various test sequences reveal a considerable improvement in the object segmentation performance when textural information is embedded. Introduction: Recent developments have principally focused upon automatic object segmentation strategies [1-4] because of their diverse application base. A statistical framework based on graphical model is found in [1], while an alternative approach [2, 3] employs a Gaussian mixture model (GMM) to represent spatiotemporal features. In [2], spatio-temporal 3D volume of pixels comprising the video is modelled by a GMM, while the strategy in [3] separates objects frame by frame. However, fully automated object segmentation based on conventional low level homogeneity criteria is still in its infancy since an object typically comprises multiple colours, and/or non-homogeneous motion in different parts. Most natural surfaces exhibit texture as it is a primitive perceptual cue, which can be visually sensed, so providing the feel of an object. To successfully segment an object thus requires consideration to be given to texture, alongside other low level features, with all features having equal priority. Video object segmentation techniques typically separate objects based upon pixel-level classifications in contrast to popular textural approximation techniques normally calculated for either the entire image or a region. Recent developments such as scale invariant feature transform and gradient location and orientation histogram [5] have exhibited promising performance in describing region, though pragmatically it is infeasible to consider them with other pixel features as they are histograms generated from the region. Markov random field model [6] considers texture involving the interactions among neighbouring pixels assuming homogeneous pixel intensity and thus is unable to represent a real object. This provided the motivation to formulate a textural representation strategy that quantifies texture as a pixel feature and enables the information to be incorporated into any