Core Concepts
Our method can identify physical parameters like length, damping, and friction directly from a single video by combining neural implicit representations for appearance modeling with neural ordinary differential equations for dynamics modeling.
Abstract
The paper proposes a method that combines neural implicit representations for appearance modeling with neural ordinary differential equations (ODEs) for modeling planar physical phenomena. This allows estimating the unknown physical parameters and initial conditions of the ODE, as well as the parameters of the appearance representations, directly from a single input video.
The key highlights are:
The method can identify physical parameters like length, damping, and friction from a single video, without requiring large training datasets.
It uses a parametric dynamics model in the form of an ODE, which enables interpretable physical parameters and long-term prediction in state space.
The combination of the neural implicit representations and the ODE-based dynamics allows for photo-realistic rendering of novel scenes with modified physical parameters.
Contrary to existing learning-based approaches, the proposed method is a per-scene model that does not suffer from generalization issues to out-of-distribution data.
The paper evaluates the method on synthetic and real-world datasets, including a pendulum, a sliding block, and a thrown ball. The results show that the method can accurately recover the physical parameters and produce high-quality predictions, outperforming baseline approaches that require large training datasets.
Stats
The length of the pendulum is recovered with a relative error of less than 4.1%.
The angle of the inclined plane for the sliding block is recovered with a relative error of 3.6%.
Quotes
"Contrary to existing learning-based approaches that require large corpora of training data, we propose a per-scene model, so that only a single short video clip that depicts the physical phenomenon is necessary."
"The unique combination of powerful neural implicit representations with rich physical models allows to synthesize high-resolution and photo-realistic imagery. Moreover, it enables physical editing by rendering novel scenes with modified physical parameters."