AltNeRF: Learning Robust Neural Radiance Field via Alternating Depth-Pose Optimization

Authors: Kun Wang, Zhiqiang Yan, Huang Tian, Zhenyu Zhang, Xiang Li, Jun Li, Jian Yang

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments showcase the compelling capabilities of Alt Ne RF in generating high-fidelity and robust novel views that closely resemble reality.
Researcher Affiliation Academia Kun Wang1, Zhiqiang Yan1, Huang Tian1, Zhenyu Zhang2, Xiang Li3, Jun Li1* and Jian Yang1* 1PCA Lab, Nanjing University of Science and Technology, China 2Nanjing University, Suzhou Campus, China 3Nankai University, China
Pseudocode No The paper describes the workflow and components of the Alt Ne RF framework and alternating algorithm using text and diagrams (Figure 3), but it does not include formal pseudocode or algorithm blocks.
Open Source Code No The paper does not contain an explicit statement about releasing the source code for the described methodology, nor does it provide a link to a code repository.
Open Datasets Yes We evaluate Alt Ne RF on four datasets: LLFF (Mildenhall et al. 2019), CO3D (Reizenstein et al. 2021), Scan Net (Dai et al. 2017) and our collected dataset, named Captures.
Dataset Splits No We employ the same train/test data division as BARF, which uses the first 90% of frames for training and the remaining 10% for testing. The paper does not specify a separate validation split percentage or details.
Hardware Specification Yes Our method is trained for 150K-200K iterations according to the number of frames, which costs around 4.0-6.4 hours totally on single RTX 3090.
Software Dependencies No The paper mentions software components and architectures like "U-Net" and "Res Net-50", but it does not specify any version numbers for these or other relevant software libraries or programming languages (e.g., Python, PyTorch).
Experiment Setup Yes The γ in Eq. (11) is set to 0.08 for LLFF and CO3D, and 0.15 for Scan Net and Captures. We pretrain the SPM with a learning rate of 10 4, and fine-tune it with 5.0 10 5. For SRM, the initial learning rate for Ne RF learning is set to 10 3, and exponentially decays to 10 4 throughout the training process. The initial learning rate for pose refinement is set to 10 5, and linearly increases to 2.0 10 3 after 1K iterations before exponentially decaying to 10 5. The number of iterations Sr and Sp are set to 50K and 500, respectively, and we perform two alternations in all experiments unless otherwise specified. Our method is trained for 150K-200K iterations according to the number of frames...