PanoGRF: Generalizable Spherical Radiance Fields for Wide-baseline Panoramas
Authors: Zheng Chen, Yan-Pei Cao, Yuan-Chen Guo, Chen Wang, Ying Shan, Song-Hai Zhang
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results on multiple panoramic datasets demonstrate that Pano GRF significantly outperforms state-of-the-art generalizable view synthesis methods for wide-baseline panoramas (e.g., Omni Syn) and perspective images (e.g., IBRNet, Neu Ray). |
| Researcher Affiliation | Collaboration | Zheng Chen1 , Yan-Pei Cao2, Yuan-Chen Guo1, Chen Wang1, Ying Shan2, Song-Hai Zhang1 1Tsinghua University 2ARC Lab, Tencent PCG China |
| Pseudocode | No | The paper describes its methods verbally and with diagrams, but does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Poject Page: https://thucz.github.io/Pano GRF/. |
| Open Datasets | Yes | We conduct experiments on Matterport3D [4], Replica [28], and Residential [8]. |
| Dataset Splits | No | The paper states which datasets are used and specifies test data, but does not provide specific training/validation/test dataset splits (e.g., percentages or exact counts) for reproducibility. |
| Hardware Specification | No | The paper mentions 'A100 GPU' and 'V100 GPU' in the context of training time, but does not provide specific models, quantities, or other detailed hardware specifications for the experimental runs. |
| Software Dependencies | No | The paper mentions general components like 'Adam optimizer' and 'Res Net' with citations, but does not specify version numbers for programming languages, libraries, or other software dependencies. |
| Experiment Setup | Yes | We set Nmono = 5, Nuni = 59, and Nl = 2. D = Nmono + Nuni is 64. σ used in mono-guided depth sampling is set to 0.5 and β is set to 3. (...) Pano GRF employs the Adam optimizer [16] with an initial learning rate of 4.0e-4. (...) a batch size of 512 was used during training. (...) The learning rate is halved every 20k iterations. |