Single image 3D human pose estimation from noisy observations

Abstract

Markerless 3D human pose detection from a single image is a severely underconstrained problem because different 3D poses can have similar image projections. In order to handle this ambiguity, current approaches rely on prior shape models that can only be correctly adjusted if 2D image features are accurately detected. Unfortunately, although current 2D part detector algorithms have shown promising results, they are not yet accurate enough to guarantee a complete disambiguation of the 3D inferred shape.\ud In this paper, we introduce a novel approach for estimating 3D human pose even when observations are noisy. We propose a stochastic sampling strategy to propagate\ud the noise from the image plane to the shape space. This provides a set of ambiguous 3D shapes, which are virtually undistinguishable from their image projections. Disambiguation is then achieved by imposing kinematic constraints that guarantee the resulting pose resembles a 3D\ud human shape. We validate the method on a variety of situations in which state-of-the-art 2D detectors yield either inaccurate estimations or partly miss some of the body parts.Preprin

A Joint Model for 2D and 3D Pose Estimation from a Single Image

“…We also compare our results with [3,6,7,24]. This is just meant to be an indicative result, as the different methods are trained and evaluated differently.…”

“…This is just meant to be an indicative result, as the different methods are trained and evaluated differently. Table 3 summarizes the results using the pose error, corresponding to the "aligned error" in [24]. The two algorithms that use temporal information [3,7] are evaluated using absolute error.…”

“…Recently, the estimations of an off-the-shelf 2D detector [29] have already been used for 3D pose estimation in [24]. Yet, and in contrast to the solution we propose here, the 2D estimations are not updated while inferring the 3D shape, and thus, the final 3D pose strongly depends on the result of the 2D detector.…”

Lie Algebra-Based Kinematic Prior for 3D Human Pose Tracking

“…We show quantitative results that demonstrate a large improvement over the widely used Gaussian diffusion models. Furthermore, it is straightforward to extend existing 3D human pose estimation algorithms [11,12] to tracking using the proposed prior using stronger image features [14,15].…”

Unconstrained Monocular 3D Human Pose Estimation by Action Detection and Cross-Modality Regression Forest

“…Figure 4: Overview of the proposed framework models are also investigated in 3D human pose estimation, for instance, [5] uses a pedestrian detector with deformable body parts to estimate rough 3D poses in street scenes, [28] optimises multiple pose hypotheses from 2D DPM using inverse kinematics to estimate 3D pose in a single image. In this work, we further investigate the use of 2D DPM in a multi-action 3D HPE scenario.…”