Accurate static pose estimation combining direct regression and geodesic extrema

Abstract

Abstract-Human pose estimation in static images has received significant attention recently but the problem remains challenging. Using data acquired from a consumer depth sensor, our method combines a direct regression approach for the estimation of rigid body parts with the extraction of geodesic extrema to find extremities. We show how these approaches are complementary and present a novel approach to combine the results resulting in an improvement over the state-of-the-art. We report and compare our results a new dataset of aligned RGB-D pose sequences which we release as a benchmark for further evaluation.

Seeing and understanding people

“…Our first work with the Kinect was to apply poselets (Bourdev et al 2010) in the depth domain ). Although part detection can still be used, as in the seminal work of (Shotton et al 2011), we chose to adopt direct regression based approaches (Holt and Bowden 2012), (Holt et al 2013), the later of which combines regression of joints with the identification of geodesic extrema. As seen in Figure 3, regression works well for the torso but degrades as the degrees of freedom of the body parts increase, leading to poor hand prediction.…”

Sparse Pose Regression via Componentwise Clustering Feature Point Representation

“…Thus, feature point approaches always integrate this technique with other complementary techniques, such as regression forest based decision [8], additional feature landmark detection [13], [1], kinematic model fitting [7] or database lookup [1].…”

“…According to the methodology, research work on pose estimation via depth data can be roughly divided into four categories: the feature point approaches, the pixel classification approaches, the probabilistic graphical model approaches and others. In general, the feature point approaches [12], [1], [13], [8], [7] detect interest points of the range image ( with the extrema of geodesic distances of depth points), and then employ a database lookup scheme (e.g., [1]), classify each interesting point as a point of a body component using its local descriptors (e.g., [12]), or detect additional feature points as the secondary feature (e.g., [1], [13]) for discriminating human poses; the pixel classification approaches [14], [15], [18], [20], [11], [3], [9] determine which body component each pixel belongs to, using the so-called depth difference feature and random decision forests; the probabilistic graphical model approaches [26], [27], [6], [23] represent a human body as a set of rigid or non-rigid parts connected with pairwise constraints and formulate the localization of body parts as an optimization of the prior probability of part configurations. Different approaches fit for different applications of visual media, i.e., the codec and reconstruction of models of human body, the retrieval and analysis of human motion.…”

Real Time Fingertip Detection with Kinect Depth Image Sequences

“…In fact, for most HCI applications, the detection of key body parts is enough and the geodesic distance [7,8,9,17,18] is a promising technique. Plagemann et al [7] models the depth data as a surface mesh and proposes a graph based method for body part identification.…”

“…As the geodesic distances along surface mesh remain constant regardless of deformation, the body parts are hypothesized as the geodesic extremas. Apart from body part identification [7,17,18], the concept of geodesic distance is also applied in gesture recognition [8,9]. Liang et al [8] identifies fingertips as points with comparatively large geodesic distances.…”