# SMPLR: Deep SMPL reverse for 3D human pose and shape recovery

**Authors:** Meysam Madadi, Hugo Bertiche, Sergio Escalera  
**arXiv:** [1812.10766](https://arxiv.org/abs/1812.10766)  
**Date:** 2018-12-27  
**Summary page:** https://www.summarizepaper.com/en/paper/1812.10766v2/

## Key points

- Significant advancements in 3D human pose and shape recovery using deep neural networks and statistical morphable body models like SMPL
- Introduction of SMPLR method to address issues with SMPL-based solutions, involving embedding SMPL within a deep model for accurate 3D pose and shape estimation from single RGB images
- Use of CNN-based 3D joint predictions as an intermediate representation similar to an autoencoder
- Advantage of SMPLR method in eliminating complex constraints on pose and shape compared to traditional approaches
- Introduction of denoising autoencoder component for datasets lacking accurate 3D annotations, lifting 2D joints to 3D without paired annotations
- Significant improvements over existing methods shown in experiments on SURREAL and Human3.6M datasets with error reductions of approximately 4 and 25 millimeters respectively
- End-to-end training approach applied by initially training all networks independently before fine-tuning collectively, with ablation study conducted to analyze effects of different module combinations
- UP-3D dataset comprising labeled images from various sources fitted with a gender-neutral SMPL model, while SURREAL dataset consisted of synthetic images generated with realistic poses under diverse conditions
- Promising results demonstrated by the proposed SMPLR method in improving accuracy and efficiency in 3D human pose and shape recovery tasks compared to existing techniques

## Summary

, , , , 

The field of 3D human pose and shape recovery has seen significant advancements with the use of deep neural networks and statistical morphable body models like the Skinned Multi-Person Linear model (SMPL). However, these SMPL-based solutions often struggle to accurately predict 3D bodies due to their unconstrained nature, resulting in unrealistic body meshes. To address this issue, a new approach called SMPLR has been proposed in this paper. The SMPLR method involves embedding SMPL within a deep model to estimate 3D pose and shape from a single RGB image, using CNN-based 3D joint predictions as an intermediate representation. This process can be likened to an autoencoder, with the encoder being a deep neural network and the decoder being the SMPL model. One key advantage of the SMPLR method is its ability to eliminate complex constraints on pose and shape that are typically required in traditional SMPL-based approaches. Additionally, for datasets lacking accurate 3D annotations, a denoising autoencoder (DAE) component has been introduced to lift 2D joints to 3D without paired annotations. Experiments on popular datasets SURREAL and Human3.6M showed significant improvements over existing methods, with error reductions of approximately 4 and 25 millimeters respectively. During training, an end-to-end approach was applied by initially training all networks independently (SHN, DAE, Ω and Ψ) before fine-tuning the entire network collectively. An ablation study was also conducted to analyze the individual effects of different combinations of modules in training. The UP-3D dataset comprised labeled images from various sources such as LSP, LSP-extended, and MPII-HumanPose datasets after fitting a gender-neutral SMPL model into them. The SURREAL dataset consisted of synthetic images of humans generated with realistic poses under diverse conditions. Overall, the proposed SMPLR method demonstrates promising results in improving accuracy and efficiency in 3D human pose and shape recovery tasks compared to existing techniques.