Mask-Based Modeling for Neural Radiance Fields
Authors: Ganlin Yang, Guoqiang Wei, Zhizheng Zhang, Yan Lu, Dong Liu
ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate the effectiveness of our proposed MRVM-Ne RF on both synthetic and real-world datasets, qualitatively and quantitatively. Besides, we also conduct experiments to show the compatibility of our proposed method with various backbones and its superiority under few-shot cases. |
| Researcher Affiliation | Collaboration | Ganlin Yang 1 Guoqiang Wei 2 Zhizheng Zhang 3 Yan Lu 3 Dong Liu 1 1 University of Science and Technology of China 2 Byte Dance Research 3 Microsoft Research Asia ygl666@mail.ustc.edu.cn weiguoqiang.9@bytedance.com {zhizzhang,yanlu}@microsoft.com dongeliu@ustc.edu.cn |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Our codes are available at https://github.com/Ganlin-Yang/MRVM-Ne RF. |
| Open Datasets | Yes | Specifically, we adopt transformer-based backbone under synthetic NMR Shape Net dataset (Kato et al., 2018), which is introduced in Section 4.1. We also employ MLP-based backbone under realistic complex scenes, with Ne RF Synthetic (Niemeyer et al., 2020), DTU (Jensen et al., 2014) and LLFF (Mildenhall et al., 2019) as the three evaluation datasets, presented in Section 4.2. |
| Dataset Splits | No | The paper mentions training, finetuning, and evaluation phases, along with various views used as references. For instance, 'The default configuration in Table 3 uses 100 views for finetuning and renders each image from 8 reference views.' However, it does not explicitly define or specify a 'validation' dataset split with percentages, sample counts, or a dedicated section for it. |
| Hardware Specification | Yes | The two-stage training takes about 10 days on GTX-1080Ti. (...) All the training processes are conducted on one V100 GPU, which takes about 5 days for total pretraining and finetuning. |
| Software Dependencies | No | All the models are implemented using Pytorch (Paszke et al., 2019) framework. |
| Experiment Setup | Yes | We sample 64 points along each ray at coarse stage, and extra 32 points at fine stage. The moving average decay rate τ in Equation 5 is set to 0.99, the default mask ratio η is set to 50% and the coefficient λ for loss term Lmrvm is set to 0.1 during mask pretraining stage unless otherwise stated. |