PC-HMR: Pose Calibration for 3D Human Mesh Recovery from 2D Images/Videos
Authors: Tianyu Luan, Yali Wang, Junhao Zhang, Zhe Wang, Zhipeng Zhou, Yu Qiao2269-2276
AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We perform extensive experiments on the popular benchmarks, i.e., Human3.6M, 3DPW and SURREAL, where our PC-HMR frameworks achieve the SOTA results. |
| Researcher Affiliation | Academia | 1Shen Zhen Key Lab of Computer Vision and Pattern Recognition, Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Shenzhen, China 2SIAT Branch, Shenzhen Institute of Artificial Intelligence and Robotics for Society 3University of California, Irvine |
| Pseudocode | No | The paper does not contain any explicitly labeled 'Pseudocode' or 'Algorithm' blocks with structured steps. |
| Open Source Code | No | All the codes and models will be released afterwards. |
| Open Datasets | Yes | To evaluate our PC-HMR frameworks, we investigate extensive experiments on three popular benchmarks, i.e., Human3.6M (Ionescu et al. 2014), 3DPW (von Marcard et al. 2018) and SURREAL (Varol et al. 2017). |
| Dataset Splits | Yes | Specifically, Human3.6M is a popular motion capture dataset. We use 5 subjects (S1, S5, S6, S7 and S8) for training and 2 subjects (S9 and S11) for testing. 3DPW is an in-the-wild dataset with multiple actors occurred in the same image. We use its official data split for training and testing. SURREAL is a large-scale synthetic dataset with SMPL body annotations. We directly evaluate its test set by our model pretrained on Human3.6M to show generalization capacity. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., GPU models, CPU types, or memory) used for running its experiments. It only mentions implementing frameworks using PyTorch. |
| Software Dependencies | No | We implement our frameworks by Py Torch. |
| Experiment Setup | Yes | In our experiments, we first pretrain HMR (including mesh-to-pose and 3D-to-2D projectors) and 2D-to-3D pose lifter separately. Then, we fine-tune the entire framework, by adding 3D pose supervision on pose lifter and 3D mesh supervision on pose calibration module. As a result, our serial PC-HMR can effectively calibrate HMR mesh by 3D pose refinement. ... In our experiments, we first pretrain mesh stream (HMR+Mesh-to Pose Projector) and pose stream separately, and then finetune the entire framework by adding 3D pose supervision on pose lifter and 3D mesh supervision on calibration module. ... we set 90/50 training epochs with 1024/128 mini-batch size for Human3.6M and 3DPW. We set the learning rate as 0.001 for our calibration module and 1 10 5 for other modules. |