Differentiable Acoustic Path Tracing: Geometry Optimization with Reparameterized Path Replay Backpropagation (en)
* Presenting author
Abstract:
Path tracing is a well-established method for basic forward acoustic simulations. Differentiating this process poses a significantly more complex challenge. In this study, we extend Mitsuba 3, a framework for differentiable light transport optimized for efficient GPU computation, to simulate sound propagation. Backpropagation of derivatives through the path tracing algorithm allows for the application of gradient descent techniques in the reconstruction of scene properties, including the optimization of absorption and scattering factors, as well as the scene geometry.To streamline the optimization process and circumvent the computational overhead associated with automatic differentiation of extensive computation graphs, we employ the technique of Path Replay Backpropagation.Furthermore, to address the discontinuities which can arise in the Monte Carlo estimation of the fundamental path tracing integral during the calculation of derivatives, particularly those related to vertex positions of geometry, we apply a reparametrization strategy using warped area sampling.We illustrate our acoustic extension's capabilities by optimizing the scene geometry in a typical room acoustic scenario.