Motion Decoupled 3D Gaussian Splatting for Dynamic Object Representation

Authors: Xiao Hu, Libo Long, Jochen Lang

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

Reproducibility Variable Result LLM Response
Research Type Experimental 4 Experimental Evaluation We select five state-of-the-art 3D representation methods as comparators. ... The evaluation metrics follow the previous public benchmarks (Pumarola et al. 2021; Li et al. 2021). Specifically, Peak Signal-to-Noise Ratio (PSNR), Structural Similarity (SSIM), and VGG-based Learned Perceptual Image Patch Similarity (LPIPS) (Zhang et al. 2018) are used.
Researcher Affiliation Academia Xiao Hu, Libo Long, Jochen Lang University of Ottawa, Canada, EMAIL
Pseudocode No No explicit pseudocode or algorithm blocks are present in the paper. The methodology is described in prose and mathematical formulas.
Open Source Code Yes Code https://github.com/haliphinx/M5D-GS
Open Datasets Yes Both the dataset with a total of ten scenes and the source files used for its creation are available open-source, allowing the community to further investigate severe motion understanding. Current public datasets (Pumarola et al. 2021; Li et al. 2021; Yan, Li, and Lee 2023) for dynamic scene representation usually contain only slight motion and deformation.
Dataset Splits No The paper introduces a novel dataset and augments existing ones, but it does not specify explicit training/validation/test splits (e.g., percentages, sample counts, or specific files) for its experiments within the main text.
Hardware Specification No The paper does not provide specific details about the hardware used to run the experiments, such as GPU or CPU models.
Software Dependencies No The paper does not provide specific software dependencies with version numbers needed to replicate the experiment.
Experiment Setup No The main constraints for the proposed M5D-GS still follow the original 3D-GS without additional loss for motion estimation. The overall constraints include a per-pixel L1 loss and a D-SSIM loss (Kerbl et al. 2023) LD SSIM. The loss function is Limg = L1 + λLD SSIM with λ as the loss coefficient. More details are available in the supplementary material.