Fine-Grained Multi-View Hand Reconstruction Using Inverse Rendering

Authors: Qijun Gan, Wentong Li, Jinwei Ren, Jianke Zhu

AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct the comprehensive experiments on Inter Hand2.6M, Deep Hand Mesh and dataset collected by ourself, whose promising results show that our proposed approach outperforms the state-of-the-art methods on both reconstruction accuracy and rendering quality.
Researcher Affiliation Academia Qijun Gan, Wentong Li, Jinwei Ren, Jianke Zhu* College of Computer Science and Technology, Zhejiang University, China {ganqijun,liwentong,zijinxuxu,jkzhu}@zju.edu.cn
Pseudocode No The paper describes its methodology in narrative text and uses equations, but it does not include any clearly labeled pseudocode or algorithm blocks.
Open Source Code Yes Code and dataset are publicly available at https://github.com/agn Jason/FMHR.
Open Datasets Yes Inter Hand2.6M. Inter Hand2.6M (Moon et al. 2020) is a large-scale dataset... Deep Hand Mesh. The Deep Hand Mesh dataset (Moon, Shiratori, and Lee 2020)... Code and dataset are publicly available at https://github.com/agn Jason/FMHR.
Dataset Splits No The paper mentions using Inter Hand2.6M for pre-training and fine-tuning, and evaluating on 'the rest views' for one specific experiment, but it does not provide comprehensive training/validation/test dataset splits (e.g., percentages or sample counts) for reproducibility across all datasets.
Hardware Specification Yes Notably, the entire optimization pipeline is computationally efficient, which takes approximately 90 seconds on a single NVIDIA 3090Ti GPU.
Software Dependencies No The paper mentions the Adam optimizer but does not list specific software dependencies (e.g., programming languages, libraries, frameworks) with their version numbers required for reproducibility.
Experiment Setup Yes During the optimization process, we utilize the Adam optimizer (Kingma and Ba 2014) with the balanced weights of λ1 = 20, λ2 = 40, λ3 = 20, λ4 = 100, and λ5 = 2 to jointly optimize the vertices, vertex albedo, and lighting coefficients over 100 iterations. Subsequently, for fine-tuning and joint optimization, each process requires 100 epochs of training with γ1 = 100 and γ2 = 2, respectively.