Sensing Surface Patches in Volume Rendering for Inferring Signed Distance Functions

Authors: Sijia Jiang, Tong Wu, Jing Hua, Zhizhong Han

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

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate our method by numerical and visual comparisons on scene benchmarks. Our superiority over the latest methods justifies our effectiveness.
Researcher Affiliation Academia Sijia Jiang, Tong Wu, Jing Hua*, Zhizhong Han Department of Computer Science, Wayne State University, Detroit, MI, USA EMAIL, EMAIL, EMAIL, EMAIL
Pseudocode No The paper describes the methodology through text and mathematical equations, but it does not include any explicitly labeled pseudocode or algorithm blocks.
Open Source Code No The paper does not explicitly state that source code is provided or offer a link to a code repository.
Open Datasets Yes Datasets. We evaluate our method by comparisons with the latest methods for scene reconstruction from multi-view images under both synthetic scenes and real scans. The synthetic indoor scenes are Replica (Straub et al. 2019), released by Mono SDF (Yu et al. 2022), the real scans of indoor scenes are Scan Net (Dai et al. 2017), released by either Mono SDF or Neual RGBD (Azinovi c et al. 2022), and real-world large-scale indoor scenes, Tanks and Temples (Knapitsch et al. 2017), released by Mono SDF.
Dataset Splits Yes Scan Net from Mono SDF. Numerical comparisons with the latest methods are presented in Tab. 1. Following Mono SDF (Yu et al. 2022), we use monocular cues estimated by the pretrained Omnidata model (Eftekhar et al. 2021). Replica. We evaluate our method on synthetic Replica scenes, following the Mono SDF setup and using monocular cues predicted by the pretrained Omnidata model. The scale and shift are solved in Tab. 1. Tab. 3 shows our superiority over Mono SDF, both with and without Replica pretraining (Train split and Test split).
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory) used for running the experiments.
Software Dependencies No The paper does not explicitly list specific software dependencies with version numbers needed to replicate the experiment.
Experiment Setup Yes Details. For each posed view, we sample 1024 rays per training batch. Using Vol SDF s error-bounded sampling strategy and architecture, we sample points along rays. To create surface patches, geometry cues are backprojected (either predicted monocular or dataset-provided sensor cues) to obtain 3D anchor points q. Around each anchor, we define an isotropic Gaussian distribution N(q, τ 2) and sample J = 9 points, with τ 2 controlling patch size. Ablation studies in Tab. 5 analyze the effect of different τ 2 values. Loss weights are set as λ1 = 0.1, λ2 = 0.05, λ3 = 0.05, λ4 = 0.5, and λ6 = 0.5 for balanced contributions. To improve coarse shape reconstruction, we apply inverse weight annealing for LNCC, setting λ5 = 0 for the first 100 epochs and gradually increasing it to 0.1.