A General Implicit Framework for Fast NeRF Composition and Rendering

Authors: Xinyu Gao, Ziyi Yang, Yunlu Zhao, Yuxiang Sun, Xiaogang Jin, Changqing Zou

AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we first demonstrate that our pipeline indeed takes effect for various Ne RF works. Following that, we show how our framework can assemble a large number of neural objects to form a virtual 3D world. Furthermore, we show our approach can be combined with traditional rendering techniques to create a mixed rendering pipeline. We implement our framework in two versions: Pytorch version for a convenient reproduction and CUDA version for testing performance limitation. We conduct experiments on our new N-object dataset which has been described in section N -object dataset testing .
Researcher Affiliation Academia Xinyu Gao1, Ziyi Yang1, Yunlu Zhao1, Yuxiang Sun2, Xiaogang Jin1*, Changqing Zou1,2 1State Key Lab of CAD&CG, Zhejiang University 2Zhejiang Lab {22121052, Ingram14, yunlu.zhao}@zju.edu.cn, sunyuxiangyx@gmail.com jin@cad.zju.edu.cn, changqing.zou@zju.edu.cn
Pseudocode Yes The overview and data flow of our render framework is shown in Fig. 1, which is also described by the algorithm pseudo-codes in appendix.
Open Source Code No The paper states 'We implement our framework in two versions: Pytorch version for a convenient reproduction and CUDA version for testing performance limitation.' However, it does not provide an explicit statement about open-sourcing the code or a link to a repository.
Open Datasets No The paper states 'First, we collected an Nobject dataset consisting of 22 distinct Ne RF objects: 7 objects were chosen from previous work (3 from Neus dataset of real-world objects, 4 from synthesis dataset (Mildenhall et al. 2020; Zhang et al. 2022; Verbin et al. 2022)), and the remaining 15 objects were newly created by Blender 3D.' While some components are from cited works, the full 'N-object dataset' collected and created by the authors is not provided with public access details (link, DOI, specific repository).
Dataset Splits No The paper mentions 'We use the pre-trained Ne RF models to generate 500 random views (more random views, less artifacts) for supervision for each Ne DF model of a single object, and train the intersection network over 60W iterations until convergence.' However, it does not explicitly provide details about standard train/validation/test dataset splits (percentages, counts, or predefined splits) for the main experimental evaluation.
Hardware Specification Yes A single Nvidia A100 GPU is used to train Ne DF for each scene from our N-object dataset. ... CUDA version of the proposed framework is built with Vulkan, Tensor RT, and a customized CUDA kernel to enable real-time manipulation on Nvidia RTX 4090 GPU.
Software Dependencies No The paper mentions 'Pytorch version', 'CUDA version', 'Vulkan', and 'Tensor RT' as software components used in their implementation but does not specify their version numbers.
Experiment Setup Yes We employ Adam as the optimizer and set the learning rate to 5e-4. We use the pre-trained Ne RF models to generate 500 random views (more random views, less artifacts) for supervision for each Ne DF model of a single object, and train the intersection network over 60W iterations until convergence. The batchsize of training rays is set to 4,096 during each iteration.