Complexity-Aware Dynamic Gradient Shifting: A Novel Soft Supervision Training Strategy for 3D Pose Estimation and Regression Learning

Recent state-of-the-art (SOTA) techniques have demonstrated substantial efficacy in 3D Human Pose Estimation (HPE) from videos. Despite strong progress, no prior work has used soft supervision to handle Hard-to-Estimate (H2E) samples in 3D pose estimation, which is inherently a regression task. Existing H2E-example mining solutions, based on logit distillation, progressive target refinement, and label smoothing, are confined to classification problems. Traditional regression-based problems are deprived of the benefits of these regularization techniques. A soft supervision-based H2E example mining technique is crucial for regression problems. <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">To the best of the author’s knowledge, there are no soft supervision-based regularization techniques exist for regression problems</i>. This paper proposes a novel training strategy, referred to as Progressive Soft-Supervision Training for Regression Problems (PSTR). PSTR introduces the concept of progressive <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">soft targets</i> to the 3D pose estimator, a regression-based task. Highly inaccurate predictions, representing the H2E poses, are focused more while preserving the representations for Easy-to-Estimate (E2E) poses. PSTR forces the network to learn an alternate optimum inductive bias in the solution space <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">via dynamically shifting gradients</i>. The proposed PSTR improves SOTA performances on the large-scale Human3.6m dataset by a large margin with an average MPJPE, and P-MPJPE of 1.2<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">mm</i> and 1.09<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">mm</i> for Protocol 1 and 2, respectively, where improvements on PCK, AUC, and MPJPE for MPI_INF_3DHP dataset are, 1.53%, 1.75% and 1.40<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">mm</i> for 2D-3D pose uplifting and 1.60%, 1.95% and 1.60<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">mm</i> for RGB to 3D pose estimation tasks, respectively. PSTR can be effortlessly deployed on any regression-based task.