Time-derivative rendering
Forward-mode rasterization produces log-luminance and its instantaneous derivative in the same visibility ordering.
TIDES: Time-Derivative Event Simulation via Deformable Reconstruction
University of Surrey, Guildford, Surrey, UK
ECCV 2026
* Equal contribution
Event stream comparison showing ground truth, TIDES, and baseline simulators.
Overview
Event cameras emit asynchronous events in response to changes in scene appearance. TIDES simulates these events directly from a dynamic Gaussian splatting scene, deriving per-pixel intensity dynamics from the scene rather than differencing rendered frames.
This continuous-time formulation predicts threshold crossings, including multiple crossings per rendering step, without temporal upsampling or frame interpolation. It is designed to avoid timestamp batching, where many events are forced to share the same discrete rendering times under fast motion or occlusion.
The simulator also uses scene visibility cues to adapt its time stepping and models finite sensor bandwidth with a tile-level readout arbiter, reproducing burst-induced timestamp spreading and event drops.
Risk-guided stepping
Adaptive sampling concentrates computation near mixed visibility, disocclusions, and fast-changing opacity.
Sensor-aware readout
A finite-bandwidth arbiter captures localized event bursts, jitter, and drops under high activity.
Supplementary Videos
The videos compare synchronized ground-truth events, simulated events, and RGB motion cues across paired event-RGB datasets. Motion reveals timing drift, burst fragmentation, polarity flicker, and short-lived edge events that can be hidden in static event frames.
EDS dark peanuts
Event stream fidelity under ego-motion and low-light appearance changes.
DSEC Zurich City 04b
Street-scene ego-motion with structured event bursts and visibility changes.
HS-ERGB spinning umbrella
Fast dynamic motion where temporal dispersion along boundaries matters.
ESRGB horse
Dynamic object motion with synchronized event and RGB reference cues.
Novel-View Stress Test
TIDES also probes viewpoint extrapolation by training from a monocular capture while the camera moves around a playing video display, then simulating events from novel views.
Novel-view events
Polarity-colored events rendered from the extrapolated trajectory.
Novel-view RGB poses
RGB reference frames for the novel camera path.
Static 4DGS poses
Scene reconstruction cues used by the deformable representation.
Paper
Christopher Thirgood, Dipon Kumar Ghosh, and Simon Hadfield. TIDES: Time-Derivative Event Simulation via Deformable Reconstruction. ECCV 2026.
BibTeX
@inproceedings{thirgood2026tides,
title = {TIDES: Time-Derivative Event Simulation via Deformable Reconstruction},
author = {Thirgood, Christopher and Ghosh, Dipon Kumar and Hadfield, Simon},
booktitle = {European Conference on Computer Vision (ECCV 2026)},
year = {2026},
url = {https://github.com/ThirgoodC/TIDES}
}