USB-NeRF: Unrolling Shutter Bundle Adjusted Neural Radiance Fields

Authors: Moyang Li, Peng Wang, Lingzhe Zhao, Bangyan Liao, Peidong Liu

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experimental evaluations are conducted with both synthetic and real datasets, to evaluate the performance of our method. The experimental results demonstrate that USB-Ne RF achieves superior performance compared to prior state-of-the-art methods (e.g. as shown in Figure 1) in terms of rolling shutter effect removal, novel view image synthesis as well as camera motion estimation.
Researcher Affiliation Academia Moyang Li1,2 Peng Wang1,3 Lingzhe Zhao1 Bangyan Liao1,3 Peidong Liu1 1Westlake University 2ETH Z urich 3Zhejiang University moyali@ethz.ch, {wangpeng, zhaolingzhe, liaobangyan, liupeidong}@westlake.edu.cn
Pseudocode No The paper describes the method details in prose (Section 3) but does not include any pseudocode or algorithm blocks.
Open Source Code Yes Code and data are available at https://github.com/WU-CVGL/USB-Ne RF.
Open Datasets Yes We also evaluate our method on a public synthetic dataset (i.e. Carla-RS (Liu et al., 2020)) for fair comparisons against other methods.
Dataset Splits No The paper describes the total number of images generated for synthetic datasets and mentions using '6 additional views' for novel view synthesis evaluation, but does not specify explicit train/validation/test dataset splits (e.g., percentages or counts) or reference predefined splits with citations for reproduction.
Hardware Specification Yes we run a total of 200K steps on an NVIDIA RTX 3090 GPU.
Software Dependencies No The paper states 'We implement our method in Py Torch' and 'We use Adam (Kingma & Ba, 2014) optimizer', but does not provide specific version numbers for PyTorch or any other software dependencies.
Experiment Setup Yes We set the learning rates to be 5 10 4 and 1 10 3 for the Ne RF and pose optimizers respectively. Both learning rates gradually decay to 5 10 5 and 1 10 5 respectively. During each training step, we randomly select 7200 pixels from all training images to minimize the loss function presented in Eq. 10 and we run a total of 200K steps on an NVIDIA RTX 3090 GPU. We adopt the linear adjustment of the positional encoding starting from steps 20K to 100K to achieve the coarse-to-fine training strategy as in BARF (Lin et al., 2021).