NeuPhysics: Editable Neural Geometry and Physics from Monocular Videos
Authors: Yi-Ling Qiao, Alexander Gao, Ming Lin
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments show that our method achieves superior mesh and video reconstruction of dynamic scenes compared to competing Neural Field approaches, and we provide extensive examples which demonstrate its ability to extract useful 3D representations from videos captured with consumer-grade cameras. |
| Researcher Affiliation | Academia | Yi-Ling Qiao Alexander Gao Ming C. Lin University of Maryland, College Park |
| Pseudocode | No | The paper describes the methodology with text and diagrams but does not include structured pseudocode or a formally labeled algorithm block. |
| Open Source Code | Yes | Our code and data are available here: https://sites.google.com/view/neuphysics |
| Open Datasets | No | For the training data (used with consent), we note that the background must contain sufficient detail, or the Structure from Motion implementation by COLMAP [51, 52] could fail. The paper mentions using 'real-world and synthetic videos' but does not provide specific access information (link, DOI, formal citation) for the datasets used for training. |
| Dataset Splits | No | The paper describes a sequential training strategy and uses novel view synthesis for evaluation, but does not explicitly mention a distinct validation set or its split details for hyperparameter tuning. |
| Hardware Specification | Yes | A full training cycle typically takes around 18 hours on an NVIDIA A5000 GPU. |
| Software Dependencies | No | The paper mentions 'Diff PD [12]' and 'Adam [27] optimizer' but does not provide specific version numbers for these or any other software components used in the experiments. |
| Experiment Setup | Yes | By default, we set w Eik = 0.1, woff = 50.0, and wdiv = 2.0, though these weights may be further tuned depending on the scene. The physics loss is minimized for 100 epochs (as described in Section 3.2) using Adam [27] optimizer with a learning rate of 0.01. We initialize Young s modulus to 2 105Pa and acceleration to 0m/s2 |