Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..

Dynamic Gaussian Splatting from Defocused and Motion-blurred Monocular Videos

Authors: Xuankai Zhang, Junjin Xiao, Qing Zhang

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments show that our method outperforms the state-of-the-art methods in generating photorealistic novel view synthesis from defocused and motion-blurred monocular videos. Our code is available at https://github.com/hhhddddddd/dydeblur. (From Abstract)
Researcher Affiliation Academia 1Sun Yat-sen University EMAIL,EMAIL, EMAIL
Pseudocode No The paper describes the methodology in text and uses diagrams (e.g., Figure 2 for the overview of the method) but does not contain explicit pseudocode or algorithm blocks.
Open Source Code No Answer: [No] Justification: The current code implementation requires additional refactoring for optimal readability. A refined open-source release is planned following code reorganization. (From NeurIPS Paper Checklist, Section 5)
Open Datasets Yes We evaluate our method on two datasets, including the one from D2RF [32] of defocus blur and the other one from Dy Blu RF [47] of motion blur.
Dataset Splits Yes We train and evaluate our method on all sequences from the two datasets, using their left-view blurred sequences for training and the corresponding right-view sharp sequences for evaluation. Note that, similar to previous methods, the downsampled images in the two datasets are utilized for training and evaluation.
Hardware Specification Yes Training on a sequence of 512 288 resolution takes approximately 1 hour on an NVIDIA RTX 3090 GPU, with a rendering speed of 65.143 fps for the same resolution.
Software Dependencies No We employ the Adam optimizer [17] to jointly optimize the Gaussians and SE(3) motion bases and the BP-Net. The learning rates are set to 1.6 10 4 for motion bases, and 5 10 4 for BP-Net. The learning rate for the Gaussians is consistent with that of the original 3DGS [15]. (From Section 4: Implementation Details) This section lists software components like 'Adam optimizer' and 'Py Torch' (from Appendix A) but does not provide specific version numbers for these dependencies.
Experiment Setup Yes We set the number of motion bases Nb to 20, and the blur kernel size K to 9. We employ the Adam optimizer [17] to jointly optimize the Gaussians and SE(3) motion bases and the BP-Net. The learning rates are set to 1.6 10 4 for motion bases, and 5 10 4 for BP-Net. The learning rate for the Gaussians is consistent with that of the original 3DGS [15]. We train each scene for 40,000 iterations and introduce unseen view information to constrain the scene starting from iteration 3,000. We set the iteration interval Nu for using unseen view information to 5. Dynamic Gaussians densification is performed at Nd = 2,500 iterations. We introduce the unified blur synthesis model at iteration 3,500 and incorporate the blur-aware sparsity constraint at Nspa = 5,500 iterations.