CRAYM: Neural Field Optimization via Camera RAY Matching

Authors: Liqiang Lin, Wenpeng Wu, Chi-Wing Fu, Hao Zhang, Hui Huang

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

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate our method on both the synthetic objects from Ne RF-Synthetic [26] and the real scenes from Urban Scene3D [22], for novel view synthesis and 3D geometry reconstruction, over denseand sparse-view settings. Compared to state-of-the-art alternatives, CRAYM produces superior results especially over fine details.
Researcher Affiliation Academia 1College of Computer Science and Software Engineering, Shenzhen University 2Department of Computer Science and Engineering, The Chinese University of Hong Kong 3School of Computing Science, Simon Fraser University linliqiang2020@gmail.com wenpengggg@gmail.com cwfu@cse.cuhk.edu.hk haoz@sfu.ca hhzhiyan@gmail.com
Pseudocode No No section or figure explicitly labeled "Pseudocode" or "Algorithm" was found.
Open Source Code No Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [No] Justification: All the experiments use open access dataset. The code is not released yet.
Open Datasets Yes We evaluate our method on the Ne RF-Synthetic dataset [26] with eight synthetic objects (Section 4.2), the LLFF dataset [25], and the real scenes from the Urban Scene3D dataset [22] (Section 4.3).
Dataset Splits Yes Each image set contains 48 images. We use 90% of them as training set and the remaining 10% images as testing data.
Hardware Specification Yes All the experiments are conducted with an Nvidia GV100.
Software Dependencies No The paper mentions software like PyTorch and CUDA in the context of BARF [20] or similar methods and AdamW optimizer, but does not specify version numbers for these general software dependencies as part of the experimental setup.
Experiment Setup Yes The geometry network Φg is implemented as a three-layer MLP with the ReLU activation for the input and hidden layers. The texture network Φt is implemented as a four-layer MLP with the ReLU activation for the input and hidden layers. The whole network is optimized with the AdamW optimizer with a learning rate of 0.01, β = [0.9, 0.99], and ϵ = 1.0e-15. The variance [34] of the geometry network Φg is initialized as 0.3 and is optimized with a learning rate of 0.001. We adopt a warm-up training for the first 500 iterations with the Linear LR scheduler.