ND-SDF: Learning Normal Deflection Fields for High-Fidelity Indoor Reconstruction

Authors: Ziyu Tang, Weicai Ye, Yifan Wang, Di Huang, Hujun Bao, Tong He, Guofeng Zhang

ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Consistent improvements on various challenging datasets demonstrate the superiority of our method. Our method outperforms previous approaches significantly in indoor reconstruction evaluations, and this superiority is validated through extensive ablation experiments.
Researcher Affiliation Academia 1State Key Lab of CAD&CG, Zhejiang University, 2Shanghai AI Laboratory
Pseudocode No The paper describes methods in regular paragraph text and mathematical equations without explicitly labeled pseudocode or algorithm blocks.
Open Source Code No The paper provides a project page URL: https://zju3dv.github.io/nd-sdf/. However, it does not explicitly state that the source code for the described methodology is released at this link, nor is it a direct link to a code repository.
Open Datasets Yes We conducted experiments on four indoor datasets: Scan Net Dai et al. (2017), Replica Straub et al. (2019), Tanksand Temples Knapitsch et al. (2017), and Scan Net++ Yeshwanth et al. (2023).
Dataset Splits Yes We followed the splits from Mono SDF and applied the same evaluation settings.
Hardware Specification Yes All experiments were conducted on an NVIDIA TESLA A100 PCIe 40GB, with each iteration sampling 4 1024 rays, totaling 128,000 training steps.
Software Dependencies No Our method was implemented using Py Torch Paszke et al. (2019). (The paper mentions PyTorch but does not specify a version number or other versioned libraries.)
Experiment Setup Yes Our network was optimized using Adam W Loshchilov & Hutter (2017) optimizer with a learning rate of 1e-3. The weights for loss terms were: λ1 = 0.05, λ2 = 0.0005, λ3 = 0.025, λ4 = 0.05. Upon adequate initialization of the deflection field, we initiated deflection angle guided sampling, photometric optimization, and unbiased rendering. All experiments were conducted on an NVIDIA TESLA A100 PCIe 40GB, with each iteration sampling 4 1024 rays, totaling 128,000 training steps. The training time was about 14 GPU hours. Our hash encoding resolution spanned from 25 to 211 across 16 levels, with the initial activation level set to 8 and activation steps set to 2000.