# Heuristic Weakly Supervised 3D Human Pose Estimation

**Authors:** Shuangjun Liu, Michael Wan, Sarah Ostadabbas  
**arXiv:** [2105.10996](https://arxiv.org/abs/2105.10996)  
**Date:** 2021-05-23  
**Summary page:** https://www.summarizepaper.com/en/paper/2105.10996v3/

## Key points

- Monocular 3D human pose estimation from RGB images has gained attention
- Existing models rely on supervised training with 3D pose ground truth data or known pose priors
- Collection of 3D pose data requires expensive motion capture devices, limiting applicability
- Authors propose a heuristic weakly supervised 3D human pose (HW-HuP) solution
- HW-HuP estimates 3D poses without ground truth data by leveraging partial pose priors and utilizing easily accessible observations
- Depth data is used for weak supervision during training but not during inference
- HW-HuP improves upon state-of-the-art models in challenging scenarios: estimating human poses in bed and infant poses in real-world environments
- HW-HuP achieves comparable performance to cutting-edge models on public benchmarks even without extensive 3D pose data

## Summary

The field of monocular 3D human pose estimation from RGB images has gained significant attention in recent years. Existing models heavily rely on supervised training with 3D pose ground truth data or known pose priors specific to their target domains. The collection of 3D pose data typically requires expensive motion capture devices, which limits the applicability of these models. To address this limitation, the authors propose a heuristic weakly supervised 3D human pose (HW-HuP) solution that can estimate 3D poses even when no ground truth 3D pose data is available. The HW-HuP approach leverages partial pose priors learned from existing 3D human pose datasets and utilizes easily accessible observations from the target domain to estimate both the 3D human pose and shape through an optimization and regression cycle. During training, depth data is employed for weak supervision but not used during inference. The authors demonstrate that HW-HuP significantly improves upon state-of-the-art models in two practical scenarios where obtaining accurate 3D pose data is challenging: estimating human poses in bed and infant poses in real-world environments. Furthermore, they show that HW-HuP achieves comparable performance to cutting-edge models on public benchmarks even when those models are trained using abundant 3D pose data. This highlights the effectiveness and robustness of the proposed approach. In summary, this paper presents a novel heuristic weakly supervised method for estimating 3D human poses without relying on ground truth data. By leveraging partial pose priors and utilizing easily accessible observations, HW-HuP demonstrates improved performance in challenging scenarios where collecting accurate 3D pose data is difficult while retaining competitive performance compared to state-of-the-art models trained with extensive 3D pose datasets on standard benchmarks.