Neural Surface Reconstruction of Dynamic Scenes with Monocular RGB-D Camera

Authors: Hongrui Cai, Wanquan Feng, Xuetao Feng, Yan Wang, Juyong Zhang

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on public datasets and our collected dataset demonstrate that NDR outperforms existing monocular dynamic reconstruction methods.
Researcher Affiliation Collaboration University of Science and Technology of China Alibaba Group
Pseudocode No The paper does not contain any pseudocode or algorithm blocks.
Open Source Code Yes We upload them as a part of the supplemental material in submission. We include related URLs in the supplemental material.
Open Datasets Yes To evaluate our NDR and baseline approaches, we use 6 scenes from Deep Deform [9] dataset, 7 scenes from Killing Fusion [54] dataset, 1 scene from AMA [59] dataset and 11 scenes captured by ourselves.
Dataset Splits No The paper describes the datasets used and reports quantitative results on specific sequences, but it does not provide explicit details on training, validation, and test data splits (e.g., percentages, sample counts, or citations to predefined splits).
Hardware Specification Yes We run most of our experiments with 6 104 iterations for 12 hours on a single NVIDIA A100 40GB GPU.
Software Dependencies No The paper mentions using the ADAM optimizer, and refers to methods like RVM [38] and Mi VOS [14] for segmentation, but it does not provide specific version numbers for any software dependencies.
Experiment Setup Yes We train our neural networks using the ADAM optimizer [32] with a learning rate 5 10 4. We run most of our experiments with 6 104 iterations... we sample 2, 048 rays per batch (128 points along each ray). We first uniformly sample 64 points, and then adopt importance sampling iteratively for 4 times (16 points each iteration). On depth map, we uniformly sample 2, 048 points per batch. For coarse-to-fine training, we utilize an incremental positional encoding strategy... The weights in Eq. 6 are set as: λ1 = 0.1, λ2 = 1.0, λ3 = 0.5, λ4 = 0.1, λ5 = 0.5, λ6 = 0.1.