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..
Omnidirectional 3D Scene Reconstruction from Single Image
Authors: Ren Yang, Jiahao Li, Yan Lu
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results validate the effectiveness of Omni3D, demonstrating significantly advanced 3D reconstruction quality in the omnidirectional space, compared to previous state-of-the-art methods. |
| Researcher Affiliation | Industry | Ren Yang Microsoft Research Asia EMAIL Jiahao Li Microsoft Research Asia EMAIL Yan Lu Microsoft Research Asia EMAIL |
| Pseudocode | No | The paper describes the methodology using text and diagrams (Figure 2), but does not contain explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | No | Project page: https://omni3d-neurips.github.io. (NeurIPS Paper Checklist: We will release codes upon internal approval. The datasets we used are publicly available.) |
| Open Datasets | Yes | We quantitatively evaluate the 3D scene reconstruction quality of Omni3D on three distinct datasets: the Tanks and Temples [13], Mip-Ne RF 360 [1], and DL3DV [19] datasets. |
| Dataset Splits | Yes | For Tanks and Temples [13] and Mip-Ne RF 360 [1], we evaluate Omni3D on their whole test sets. For DL3DV [19], we randomly select test scenes non-overlapping with the training samples of MVD. In each test sample, we randomly select groundtruth views in the entire omnidirectional space. |
| Hardware Specification | Yes | We conduct experiments on a machine with 8 NVIDIA A100 GPUs. In Table 5, We further show the results on a single A100 GPU. |
| Software Dependencies | No | For the 3D reprojection operations, we utilize Py Torch3D [28], which provides differentiable rasterization and rendering functions. The paper mentions software like Py Torch3D, Lo RA-tuned Cog Video X, MASt3R, and Pn P, but does not specify version numbers for reproducibility. |
| Experiment Setup | Yes | In Omni3D, we address this trade-off by empirically setting the window size N = I/4. For our default setting where I = 48 views are generated per orbit, N is therefore set to 12. In our experiments, we empirically observed that the estimated poses consistently converge after three updates (in addition to the initial pose estimation). Consequently, we set the number of iterations to 3 in our proposed PVO method. |