Towards Robust and Expressive Whole-body Human Pose and Shape Estimation

Authors: Hui En Pang, Zhongang Cai, Lei Yang, Qingyi Tao, Zhonghua Wu, Tianwei Zhang, Ziwei Liu

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We perform comprehensive experiments to demonstrate the effectiveness of Robo SMPLX on body, hands, face and whole-body benchmarks. Codebase is available at https://github.com/robosmplx/robosmplx. and 5 Experiments
Researcher Affiliation Collaboration Hui En Pang1, Zhongang Cai1,2, Lei Yang2, Qingyi Tao2, Zhonghua Wu2, Tianwei Zhang1, Ziwei Liu1 1S-Lab, Nanyang Technological University 2Sense Time Research
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code Yes Codebase is available at https://github.com/robosmplx/robosmplx.
Open Datasets Yes For whole-body training, we employ Human3.6M (H36M) [13], COCO-Wholebody [14] (the whole-body version of MSCOCO [29]) and MPII [1]. The 3D pseudo-ground truths for training are acquired using Neural Annot [36].
Dataset Splits No The paper lists datasets used for training and evaluation but does not explicitly specify the training/validation/test split percentages, sample counts, or the methodology for creating these splits within the main text.
Hardware Specification No The paper does not specify the exact hardware (e.g., GPU/CPU models, memory) used for running the experiments.
Software Dependencies No The paper does not provide specific version numbers for software dependencies or libraries used in the implementation.
Experiment Setup Yes Subnetworks are trained separately, then integrated in a multi-stage manner. Initial whole-body training runs for 20 epochs. The hand and face modules are substituted with the trained Hand and Face subnetworks, followed by 20 epochs of fine-tuning to better unify the knowledge from the Hand and Face subnetworks into the whole-body understanding. Each subnetwork is trained by minimizing the following loss function L: L = λ3DL3D + λ2DL2D + λBMLBM + λproj Lproj + λsegm Lsegm + λcon Lcon (1)