Articulated people detection and pose estimation: Reshaping the future

Abstract

State-of-the-art methods for human detection and pose estimation require many training samples for best performance. While large, manually collected datasets exist, the captured variations w.r.t. appearance, shape and pose are often uncontrolled thus limiting the overall performance. In order to overcome this limitation we propose a new technique to extend an existing training set that allows to explicitly control pose and shape variations. For this we build on recent advances in computer graphics to generate samples with realistic appearance and background while modifying body shape and pose. We validate the effectiveness of our approach on the task of articulated human detection and articulated pose estimation. We report close to state of the art results on the popular Image Parsing [25] human pose estimation benchmark and demonstrate superior performance for articulated human detection. In addition we define a new challenge of combined articulated human detection and pose estimation in real-world scenes.

Holistic Human Pose Estimation with Regression Forests

“…We have used the set of 100 train images for our regression forest. This is significantly lower in contrast to Pischulin et al ( [2], [33]), where they train with 1000 images. Similarly, Johnson and Everingham [8] train with 10000 images.…”

A Multi-layer Composite Model for Human Pose Estimation

“…According to our understanding, Eichner et al [5] proposed variant 1B, but their publiclyreleased software toolkit implemented 2B which yields higher scores. Yang et al [24] also used 2B, while both Pischulin et al [13] and Wang et al [23] use 1A. Unfortunately, these seemingly subtle variations lead to significant differences.…”

“…We evaluate our results using the Percentage of Correct Parts (PCP) metric, which counts the fraction of body parts that are correctly localized compared to the ground-truth (within some threshold). Unfortunately, as pointed out in [13], the PCP scoring metric has been implemented in slightly different ways in different papers, which has led to some confusion in the literature. These differences fall along two different dimensions.…”

“…We follow the two latter papers and also use 1A, which we hope will become the standard definition, but also report results under the other variants to illustrate the significant differences they create. Note that [13] does not report PCP numbers for individual parts, but rather combines right and left parts together. We do the same, and also average the PCP of the left and right limbs reported by [23] to convert their results into this metric as well.…”

Poselet Conditioned Pictorial Structures

“…During test time we additionally run each detector for +/-7.5 degrees to compensate for slight rotations. Torso prediction is done using the detector from [16] that we augment with a spatial prior learned on the training set.…”