PhyRecon: Physically Plausible Neural Scene Reconstruction

Authors: Junfeng Ni, Yixin Chen, Bohan Jing, Nan Jiang, Bin Wang, Bo Dai, Puhao Li, Yixin Zhu, Song-Chun Zhu, Siyuan Huang

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments demonstrate that PHYRECON significantly improves the reconstruction quality. Our results also exhibit superior physical stability in physical simulators, with at least a 40% improvement across all datasets, paving the way for future physics-based applications.
Researcher Affiliation Collaboration 1 Tsinghua University 2 State Key Laboratory of General Artificial Intelligence, BIGAI 3 Peking University
Pseudocode Yes In Alg. 1, we illustrate the process of one training iteration of PHYRECON.
Open Source Code Yes https://phyrecon.github.io
Open Datasets Yes We conduct experiments on both the synthetic dataset Replica [62] and real datasets Scan Net [9] and Scan Net++ [76].
Dataset Splits No The paper mentions using ScanNet, ScanNet++, and Replica datasets, but does not explicitly provide the train/validation/test splits used for these datasets.
Hardware Specification Yes All experiments are conducted on a single NVIDIA-A100 GPU.
Software Dependencies No The paper mentions PyTorch, Adam optimizer, and Diff Taichi but does not specify their version numbers.
Experiment Setup Yes We implement our model in Py Torch [53] and utilize the Adam optimizer [28] with an initial learning rate of 5e 4. We sample 1024 rays per iteration. When incorporating physics-guided pixel sampling, we allocate 768 rays for physics-guided pixel sampling and the remaining 256 rays for random sampling. Our model is trained for 450 epochs on Scan Net [9] and Scan Net++ [76] datasets, and 2000 epochs on Replica [62] dataset. As introduced in Sec. 3.4, training is divided into three stages. For the Scan Net [9] and Scan Net++ [76] datasets, the second and final stages begin at the 360th and 430th epochs, respectively, while for the Replica [62] dataset, these stages start at 1700th and 1980th epochs. All experiments are conducted on a single NVIDIA-A100 GPU. Following previous work [77, 32], we set 1, 0.1, 0.05, 0.04, 0.05, 0.1, 0.1 and 0.1 as loss weights for LRGB, LD, LN, LS, LEikonal, Lbs, Lop, Lrd, respectively. Additionally, we set ΞΎ = 100 for updating Gu-phy, initialize the loss weight for Lphy as 60, and increase it by 30 per epoch.