VideoHumanMIB: Unlocking Appearance Decoupling for Video Human Motion In-betweening

Authors: Haiwei Xue, Zhensong Zhang, Minglei Li, Zonghong Dai, Fei Yu, Fei Ma, Zhiyong Wu

IJCAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct extensive experiments to validate our hypothesis on several datasets, including MHAD, NATOPS, and the additional HD video collection dataset. The empirical results and ablative studies show our method consistently achieves significant improvements over most VFI methods.
Researcher Affiliation Collaboration 1Tsinghua University 2Huawei Noah s Ark Lab 3Beijing Ruxiaoyi Intelligent Technology Co., Ltd. 4Beijing Jidian Qiyuan Info Tech Co. Ltd 5Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ)
Pseudocode No The paper describes the methodology using textual explanations, mathematical equations (e.g., in Section 3.2), and a system overview diagram (Figure 2), but it does not include any clearly labeled pseudocode or algorithm blocks.
Open Source Code No The paper does not contain any explicit statement about releasing its own source code or provide a link to a code repository for the methodology described. It references external tools like MMPose for pose estimation but not its own implementation.
Open Datasets Yes Datasets We perform experiments on the following full-body human video datasets: MHAD [Chen et al., 2015] human action dataset... NATOPS [Song et al., 2011] dataset...
Dataset Splits Yes From these recordings for each motion, one is selected as a test sample, while the remaining ones are utilized as training data.
Hardware Specification Yes Our model is implemented in Py Torch on a 40GB Nvidia RTX A100 GPU
Software Dependencies No The paper mentions that the model is implemented in PyTorch, but it does not specify a version number for PyTorch or any other software libraries or dependencies. It also refers to the RTMPose method but without specific software version details.
Experiment Setup Yes We train using the Adam optimizer with a learning rate of 1e-4, running for 800 epochs with a batch size of 64.