Citation Context

...ence can be expressed as the invariance of the labels with respect to the projection from the previous frame [30] [21], or as the fitting with a location model that depends on several previous frames =-=[31]-=-. Patras et al. [21] use a MRF framework in which the Gibbs potential of a single site (region) takes into account the number of pixels for which this region has the same label as the motioncompensate...

Citation Context

...D edges or 3D geometry models. In [5], the color histogram of a target region is used as the appearance representation. In [8], a contour-based face-and-shoulder model is used for modeling people. In =-=[18]-=-, a Gaussian distribution of the object pixel values is used. These algorithms build appearance models from examples in training datasets, and then use the model to track objects. Appearance-based rep...

Citation Context

...usions. Brostow et al. [2] automatically decompose video sequences into consitituent layers sorted by depths by combining spatial information with temporal occlusions. Jojic et al. [6] and Tao et al. =-=[14]-=- both model videos as a layered composition of objects and use EM to infer objects appearances and motions. Zhou et al. [19] extend [14] by introducing the concept of background occluding layers and e...

Citation Context

...nsor data affects the interpretation by HVS: the right image is arguably easier for pedestrian detection than the left one. representations have been widely used by object tracking in visible imagery =-=[5]-=-. For IR imagery, layered representation is also attractive because it facilitates the statistical modeling of sensor data. We propose to decompose an IR image into two layers: background (still objec...

Citation Context

...ulti-object tracking from a mobile platform is a core capability for many applications in mobile robotics and autonomous vehicles [14]. While early approaches have been developed for aerial scenarios =-=[15, 16]-=-, an application on groundlevel poses significant additional difficulties. Robust multi-person tracking in such challenging situations has only recently become feasible by the development of powerful ...

Citation Context

...t representations is vast. Some of the approaches include template-based methods [15], image-statistics-based methods [7], contour-based methods [16], feature-based methods [17], and layer-based ones =-=[18]-=-. In this paper, we use color distributions as our representation due to their simplicity, efficiency, and robustness to rotation, scaling, and partial occlusions. The early work of Swain and Ballard ...

Citation Context

...However, without any object model, it is almost impossible to avoid some drift after a few seconds of tracking. For long-term tracking, the use of appearence-based models such as templates [9], [15], =-=[16]-=- lead to more robust results. However, a template representation does not allow for large changes of appearence over time. To handle appearance changes, an often difficult template adaptation step is ...

Citation Context

... the foreground and a layer for the background. Layered representation of a video was proposed by Wang et al. [22] and extended by many other researchers, including Darrell et al. [23] and Tao et al. =-=[24]-=-. In this paper, a layer model of an object at frame t is defined as Layert = (P Ot, At, Zt), where P Ot is denoted as the location of an object in the frame. We assign this value as the geometric cen...

Citation Context

...ing models of their appearance and motion. Dense motion segmentation methods [16] use optical ow and normalized graph cut to segment a graph formed from the video. Layered motion segmentation methods =-=[22,19,25,12]-=- segment videos into segments with consistent motion based on some notion of consistency. 3 Overview We begin by introducing some notation. Our Bayesian ltering framework is represented by the graphic...