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..
TreeSplat: Mergeable Tree for Deformable Gaussian Splatting
Authors: Qiuhong Shen, Xingyi Yang, Xinchao Wang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 5 Experiments |
| Researcher Affiliation | Academia | 1National University of Singapore 2The Hong Kong Polytechnic University |
| Pseudocode | No | The paper describes the methodology in prose and mathematical formulations within Section 4 'Methodology' but does not include any explicit pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code and results are available at https://github.com/florinshen/treesplat. |
| Open Datasets | Yes | D-Ne RF Dataset [7]. The D-Ne RF dataset comprises 8 synthetic dynamic scenes... Neural 3D Video Dataset. The Neural 3D Video (N3V) dataset consists of 6 indoor dynamic scenes... |
| Dataset Splits | Yes | Following standard protocols, we evaluate on test views from novel camera positions within the same temporal range as the training data. For initialization, we uniformly sample 100,000 points within the cubic volume [ 1.2, 1.2]3. Quantitative results are reported in Tab. 1 in terms of PSNR, SSIM, and LPIPS. Our Tree Splat achieves the highest reconstruction quality with an average PSNR of 37.11 d B, while maintaining a compact model size of 28 MB and requiring only 4 minutes of training per scene. After tree merging, Tree Splat reaches a rendering speed of 230 FPS, outperforming all baselines. Moreover, it uses significantly fewer Gaussians, averaging only 85K per scene, demonstrating superior efficiency without compromising quality. |
| Hardware Specification | Yes | All experiments are conducted on an NVIDIA RTX4500 Ada GPU. |
| Software Dependencies | No | The paper mentions 'CUDA implementation' but does not provide specific version numbers for any software dependencies or libraries used for implementation. |
| Experiment Setup | Yes | To balance efficiency and performance, we model the time-varying shared basis fθ in Eq. 3 using a MLP with three hidden layers, each of width 512. The sinusoidal time embedding γ(t) is set as order L = 32 to capture high-frequency temporal variations. The number of basis vectors B is set to 10 for the D-Ne RF dataset and 16 for the Neural 3D Video dataset, to accommodate the complexity of real-world scenes. Motion tree construction begins at iteration 500 in conjunction with Gaussian densification. Depth Promotion is performed at first step, after which Leaf Expansion is applied every 100 iterations and Depth Promotion every 500 iterations, alternating after every four rounds of Leaf Expansion. Densification is halted at iteration 15,000, after which the motion tree is fixed to capture stable collaborative motion patterns among Gaussians. ... The learning rate for the decay factor β is set to 5 10 3, and the learning rate for the shared MLP fθ is set to 5 10 4. All experiments are conducted on an NVIDIA RTX4500 Ada GPU. For additional hyperparameter configurations, please refer to the configuration table provided in the Appendix. |