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..
Orientation-anchored Hyper-Gaussian for 4D Reconstruction from Casual Videos
Authors: Junyi Wu, Jiachen Tao, Haoxuan Wang, Gaowen Liu, Ramana Kompella, Yan Yan
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We validate the effectiveness of Ori GS on a range of casual monocular videos, including DAVIS [59], Open AI SORA [6], You Tube-VOS [74], and the Dy Check benchmark [19]. Ori GS consistently recovers sharper geometry and more coherent motion compared to recent state-of-the-art methods, demonstrating its superior reconstruction fidelity in real-world dynamic scenes. |
| Researcher Affiliation | Collaboration | 1University of Illinois Chicago 2Cisco Research |
| Pseudocode | No | The paper describes methods and equations but does not contain explicit pseudocode or algorithm blocks. |
| Open Source Code | Yes | https://github.com/adreamwu/Ori GS |
| Open Datasets | Yes | Our primary evaluation includes videos from DAVIS [59], Open AI SORA [6], and You Tube-VOS [74], which reflect the casual and unconstrained nature of our target setting. For quantitative benchmarking, we additionally report results on the Dy Check dataset [19], which contains seven scenes with multi-camera captures for novel view synthesis evaluation. |
| Dataset Splits | Yes | For the Dy Check dataset [19], which provides multi-camera setups for quantitative assessment, we evaluate our method using standard image quality metrics: Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM) [68], and Learned Perceptual Image Patch Similarity (LPIPS) [78]. The Dy Check dataset uses three synchronized cameras: one hand-held mobile view and two stationary references. Following [19], we measure reconstruction quality from the reference viewpoints. |
| Hardware Specification | Yes | All experiments are conducted on a single NVIDIA RTX A6000 GPU, and the full optimization of a typical scene takes approximately 0.5 2 hours, depending on video length and complexity. |
| Software Dependencies | No | The paper mentions using "3D Gaussian Splatting (3DGS) [37]" as the base representation and "Spatial Tracker [73]" and "Depth Crafter [29]" but does not provide specific version numbers for these or other software components. |
| Experiment Setup | Yes | The loss function combines: (i) photometric loss [37], an RGB reconstruction loss between rendered and ground-truth images, (ii) 2D correspondence loss, alignment to long-range 2D tracks and depth priors from foundation models, as in [64, 39, 44, 54], and (iii) deformation regularization, an as-rigid-as-possible constraint [73, 63, 1, 32] applied to anchor-guided transformations. For scalability and high-fidelity reconstruction, we also design a pruning-and-densification scheme: Gaussian primitives with low opacity are pruned, while spatial regions exhibiting high response gradients with respect to µp and µ p are densified through local duplication. All experiments are conducted on a single NVIDIA RTX A6000 GPU, and the full optimization of a typical scene takes approximately 0.5 2 hours, depending on video length and complexity. |