Streaming Radiance Fields for 3D Video Synthesis

Authors: Lingzhi LI, Zhen Shen, Zhongshu Wang, Li Shen, Ping Tan

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on challenging video sequences demonstrate that our approach is capable of achieving a training speed of 15 seconds per-frame with competitive rendering quality, which attains 1000 speedup over the state-of-the-art implicit methods.
Researcher Affiliation Industry Lingzhi Li Alibaba Group llz273714@alibaba-inc.com Zhen Shen Alibaba Group zackary.sz@alibaba-inc.com Zhongshu Wang Alibaba Group zhongshu.wzs@alibaba-inc.com Li Shen Alibaba Group lshen.lsh@gmail.com Ping Tan Alibaba Group pingtan@sfu.ca
Pseudocode No The paper does not contain any blocks explicitly labeled 'Pseudocode' or 'Algorithm'.
Open Source Code Yes Code is available at https://github.com/Algo Hunt/Stream RF.
Open Datasets Yes Neural 3D Video (N3DV) Dataset [10]... Following the setting in the original paper [10]... [10] is 'Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5521–5531, 2022.' (from bibliography).
Dataset Splits No The paper describes train/test splits for the datasets but does not explicitly mention a validation split, nor does it provide percentages or sample counts for the splits.
Hardware Specification Yes All the results are recorded with our 3090 GPU except the results of N3DV are referred to the numbers in the original paper.
Software Dependencies No The paper mentions 'RMSProp' and 'ZLIB' but does not provide specific version numbers for these software components. For instance, 'ZLIB' is cited with a year '2017' but no precise version number.
Experiment Setup Yes The total training takes 128K iterations with a batch size of 5K rays. We adopt the RMSProp [19] optimizer with a decay parameter of 0.95 for optimization. We adopt a slighter larger TV penalty to mitigate foggy issues, where λTV is set to 5 × 10−4 for opacity and 5 × 10−3 for color features.