• This is the official PyTorch implementation of Cross-Attention of Disentangled Modalities for 3D Human Mesh Recovery with Transformers (ECCV 2022).
• FastMETRO (Fast MEsh TRansfOrmer) has a novel transformer encoder-decoder architecture for 3D human pose and mesh reconstruction from a single RGB image. FastMETRO can also reconstruct other 3D objects such as 3D hand mesh.
• Compared with the encoder-based transformers (METRO and Mesh Graphormer), FastMETRO-S is about 10× smaller and 2.5× faster and FastMETRO-L is about 4× smaller and 1.2× faster in terms of transformer architectures.

Transformer encoder architectures have recently achieved state-of-the-art results on monocular 3D human mesh reconstruction, but they require a substantial number of parameters and expensive computations. Due to the large memory overhead and slow inference speed, it is difficult to deploy such models for practical use. In this paper, we propose a novel transformer encoder-decoder architecture for 3D human mesh reconstruction from a single image, called FastMETRO. We identify the performance bottleneck in the encoder-based transformers is caused by the token design which introduces high complexity interactions among input tokens. We disentangle the interactions via an encoder-decoder architecture, which allows our model to demand much fewer parameters and shorter inference time. In addition, we impose the prior knowledge of human body's morphological relationship via attention masking and mesh upsampling operations, which leads to faster convergence with higher accuracy. Our FastMETRO improves the Pareto-front of accuracy and efficiency, and clearly outperforms image-based methods on Human3.6M and 3DPW. Furthermore, we validate its generalizability on FreiHAND.

We provide two ways to install conda environments depending on CUDA versions.

Please check Installation.md for more information.

We provide guidelines to download pre-trained models and datasets.

Please check Download.md for more information.

• Model checkpoints were obtained in Conda Environment (CUDA 11.1)
• To use SMPL parameter regressor, you need to set --use_smpl_param_regressor as True

We provide guidelines to run end-to-end inference on test images.

Please check Demo.md for more information.

We provide guidelines to train and evaluate our model on Human3.6M, 3DPW and FreiHAND.

Please check Experiments.md for more information.

This repository provides several experimental results:

This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No. 2022-0-00290, Visual Intelligence for Space-Time Understanding and Generation based on Multi-layered Visual Common Sense; and No. 2019-0-01906, Artificial Intelligence Graduate School Program (POSTECH)).

Our repository is modified and adapted from these amazing repositories. If you find their work useful for your research, please also consider citing them:

• METRO
• MeshGraphormer
• Pose2Mesh
• I2L-MeshNet
• GraphCMR
• HMR
• DETR
• CoFormer

This research code is released under the MIT license. Please see LICENSE for more information.

SMPL and MANO models are subject to Software Copyright License for non-commercial scientific research purposes. Please see SMPL-Model License and MANO License for more information.

We use submodules from third party (hassony2/manopth). Please see NOTICE for more information.

Junhyeong Cho ([email protected])

FastMETRO ([email protected])

If you find our work useful for your research, please consider citing our paper:

@InProceedings{cho2022FastMETRO,
    title={Cross-Attention of Disentangled Modalities for 3D Human Mesh Recovery with Transformers},
    author={Junhyeong Cho and Kim Youwang and Tae-Hyun Oh},
    booktitle={European Conference on Computer Vision (ECCV)},
    year={2022}
}