Title: TIDES: Time-Derivative Event Simulation via Deformable Reconstruction

Event cameras generate asynchronous signals triggered by changes in the visual environment. Due to the limited availability of real-world event datasets, simulation has become a critical tool for development. However, current simulators typically derive event timestamps from sequences of static frames. This approach often results in a phenomenon we identify as "timestamp batching," where numerous threshold crossings are forced to share a limited set of discrete time points. This limitation becomes particularly pronounced during rapid movement or when occlusions occur.

To address these challenges, we introduce TIDES, a continuous-time event simulator grounded in dynamic Gaussian splatting. Unlike traditional methods, TIDES leverages an explicit 3D scene representation that captures both learned geometry and motion. By deriving per-pixel intensity dynamics directly from the scene data rather than calculating differences between rendered frames, TIDES achieves superior accuracy. It can predict threshold crossings precisely, including multiple crossings within a single rendering step, without the need for temporal upsampling or frame interpolation.

Furthermore, the underlying 3D model allows TIDES to identify regions of partial occlusion between objects. The system utilizes this information to implement adaptive time stepping, focusing computational resources only on areas where occlusion dynamics render simple brightness change models insufficient. Additionally, we incorporate a model for finite sensor bandwidth, employing a tile-level arbiter to replicate realistic sensor artifacts such as throughput limits, jitter, and event drops.

Evaluations across paired RGB-event benchmarks demonstrate that TIDES achieves state-of-the-art fidelity in event stream generation. Moreover, our results indicate that events simulated using TIDES transfer more effectively to real-world downstream tasks compared to those generated by competing methods.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC