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..
EF-3DGS: Event-Aided Free-Trajectory 3D Gaussian Splatting
Authors: Bohao Liao, Wei Zhai, Zengyu Wan, Zhixin Cheng, Wenfei Yang, Yang Cao, Tianzhu Zhang, Zheng-Jun Zha
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate our method on the public Tanks and Temples benchmark and a newly collected real-world dataset, Real Ev-DAVIS. Our method achieves up to 3d B higher PSNR and 40% lower Absolute Trajectory Error (ATE) compared to state-of-the-art methods under challenging high-speed scenarios. |
| Researcher Affiliation | Academia | 1 University of Science and Technology of China EMAIL, EMAIL |
| Pseudocode | Yes | Algorithm 1 Overall Training Pipeline. |
| Open Source Code | No | Our source code and the self-collect dataset will be openly accessible after the paper is accepted. |
| Open Datasets | Yes | The datasets that support the findings of this study are available in the following repositories: Tanks and Temples [50] at https://www.tanksandtemples.org/ under CC BY 4.0 license |
| Dataset Splits | Yes | To evaluate the robustness under varying camera speeds, we employ varying temporal downsampling of 6 FPS, 4 FPS, 3 FPS, 2 FPS, and 1 FPS. We select every ten frames as a test image for NVS evaluation following Local RF [8]. For SLOW scenarios, we retain every second frame, while for FAST scenarios, we keep only one frame per five frames. |
| Hardware Specification | Yes | We test the speed on RTX2080ti. |
| Software Dependencies | No | The paper does not provide specific software dependencies with version numbers for the paper's methodology. |
| Experiment Setup | Yes | We follow the optimization parameters by the configuration outlined in the 3DGS [4]. We optimize the camera poses in the representation of quaternion rotation. The initial learning rate is set to 10 5 and gradually decays to 10 6 until convergence. The balancing weight λcm, λgrad and λP BA is empirically set to 0.1, 0.2 and 0.5. For the division of events between adjacent frames, we maintain a constant interval of 1 6s for Tanks and Temples and 1 25s for Real Ev-DAVIS, setting the number of subinterval N accordingly. For example, in Tanks and Temples, N equals 2 for 3FPS and 6 for 1FPS. This ensures adherence to the constant brightness assumption within each sub-interval and provides adequate events for the following CMax warping. The intervals of neighboring warping r in CMax are set to 3. The contrast threshold C is set to 0.25 for Tanks and Temples and 0.21 for Real Ev-DAVIS. |