17
1

Monocular 3D Pose Recovery via Nonconvex Sparsity with Theoretical Analysis

Abstract

For recovering 3D object poses from 2D images, a prevalent method is to pre-train an over-complete dictionary D={Bi}iD\mathcal D=\{B_i\}_i^D of 3D basis poses. During testing, the detected 2D pose YY is matched to dictionary by YiMiBiY \approx \sum_i M_i B_i where {Mi}iD={ciΠRi}\{M_i\}_i^D=\{c_i \Pi R_i\}, by estimating the rotation RiR_i, projection Π\Pi and sparse combination coefficients cR+Dc \in \mathbb R_{+}^D. In this paper, we propose non-convex regularization H(c)H(c) to learn coefficients cc, including novel leaky capped 1\ell_1-norm regularization (LCNR), \begin{align*} H(c)=\alpha \sum_{i } \min(|c_i|,\tau)+ \beta \sum_{i } \max(| c_i|,\tau), \end{align*} where 0βα0\leq \beta \leq \alpha and 0<τ0<\tau is a certain threshold, so the invalid components smaller than τ\tau are composed with larger regularization and other valid components with smaller regularization. We propose a multi-stage optimizer with convex relaxation and ADMM. We prove that the estimation error L(l)\mathcal L(l) decays w.r.t. the stages ll, \begin{align*} Pr\left(\mathcal L(l) < \rho^{l-1} \mathcal L(0) + \delta \right) \geq 1- \epsilon, \end{align*} where 0<ρ<1,0<δ,0<ϵ10< \rho <1, 0<\delta, 0<\epsilon \ll 1. Experiments on large 3D human datasets like H36M are conducted to support our improvement upon previous approaches. To the best of our knowledge, this is the first theoretical analysis in this line of research, to understand how the recovery error is affected by fundamental factors, e.g. dictionary size, observation noises, optimization times. We characterize the trade-off between speed and accuracy towards real-time inference in applications.

View on arXiv
Comments on this paper