EventTracer | Fast Path Tracing-based Event Stream Rendering

Abstract

Render event streams from 3D scenes, not just from finished RGB videos.

Event cameras fire asynchronous positive and negative events whenever pixel brightness changes, making them attractive for fast motion, high dynamic range, robotics, autonomous driving, and AR/VR perception. Yet large-scale Event-RGB data remain difficult to collect because real sensors require careful hardware setup, calibration, synchronization, and task-specific annotations.

EventTracer is a path tracing-based rendering pipeline for efficient and physics-aware event simulation. It renders low-sample-per-pixel RGB sequences at high temporal resolution, then uses a lightweight event spiking network to convert noisy illuminance changes into realistic event pulses. A bipolar leaky integrate-and-fire unit and bidirectional EMD loss help model event polarity, internal state bias, and saturation behavior.

The resulting pipeline generates one second of 360p event video in about 1.12 minutes, supports 1,000 FPS event timestamps, and produces synthetic Event-RGB data that transfers better to real-world event vision tasks than common video-to-event simulators.

Key Features

01

Path-Traced Event Rendering

Low-SPP Monte Carlo path tracing keeps 3D geometry, lighting, materials, and motion in the event-generation loop.

02

Lightweight EvSNet

A pixel-wise temporal network predicts event pulses from noisy illuminance sequences and can be integrated into Falcor with TensorRT.

03

Physics-Aware Spiking

BiLIF activation models positive and negative polarity, internal state bias, and saturation effects that are hard to capture with frame differencing alone.

04

Real2Sim and Sim2Real

Evaluation combines event statistics with downstream event-to-video and event-to-depth transfer to test whether simulated data behaves like real data.

Method Overview

Use noisy high-FPS path tracing as the signal, then learn how an event sensor should fire.

EventTracer pipeline with path tracing, EvSNet training, ETScenes construction, and downstream evaluation. — EventTracer renders low-SPP and high-SPP sequences from 3D scenes, trains EvSNet with V2E-generated pseudo ground truth, and builds ETScenes for downstream validation.

Traditional event simulators often work after RGB rendering is complete, which makes temporal resolution depend on costly noiseless frames. EventTracer instead makes event generation part of the rendering process: low-SPP path tracing provides high-FPS noisy illuminance, while EvSNet decides which changes correspond to event pulses.

This design keeps the advantages of 3D graphics, including controllable camera trajectories, scene geometry, depth maps, and future annotations such as normals, optical flow, and segmentation masks.

Internal state bias and saturation effect in event sensors. — Event pixels with different internal states can fire at different timestamps under the same brightness change, and strong changes may create trailing events.

Event cameras are not simple frame-difference machines. Each pixel carries an internal state, and large illuminance changes can trigger a sequence of events rather than a single response. EventTracer uses BiLIF spiking to keep these dynamics explicit.

The network is deliberately local and lightweight, with a temporal receptive field of 43, so that it can run per pixel in parallel without reading an entire video clip.

Results

Faster rendering, more stable events, and better downstream transfer.

1.12 min per second of 360p event video

1,000 FPS event timestamp precision

0.071 event-scene alignment RMSE

0.830 pixel-wise correlation

Event stream comparisons between EventTracer, V2E, V2CE, and RGB frames. — EventTracer preserves more high-frequency scene structures than V2E and V2CE across Bistro scenes, especially where camera motion creates alternating event signals.

Event ratio histograms for EventTracer, V2E, V2CE, DVS-Voltmeter, and pseudo ground truth. — Event ratio distributions are more concentrated and closer to pseudo ground truth; the Wasserstein distance is 32.4% lower than DVS-Voltmeter and about 64% lower than V2E/V2CE.

Real2Sim event-to-video reconstruction comparisons. — In Event-to-Video reconstruction, simulated events from EventTracer retain details such as street tiles, windows, chairs, and paintings.

Real2Sim event-to-depth estimation comparisons. — In Event-to-Depth estimation, EventTracer events support more consistent layouts and clearer depth structures than V2E and V2CE.

Event simulation under fast motion with EventTracer, V2E, and V2CE. — Under 10x faster camera motion, EventTracer produces trailing event activations that reflect saturation behavior in real event sensors.

Event-scene alignment statistics over six scenes and twelve intervals per scene
Method	RMSE ↓	KL ↓	Corr ↑
EventTracer	0.071	0.833	0.830
V2E	0.112	4.657	0.472
V2CE	0.117	7.165	0.581
DVS-Voltmeter	0.098	1.266	0.711

Video

Supplementary event-rendering results.

BibTeX

Citation

@article{li2026eventtracer,
  title={EventTracer: Fast Path Tracing-based Event Stream Rendering},
  author={Li, Zhenyang and Bai, Xiaoyang and Lu, Jinfan and Shen, Pengfei and Lam, Edmund Y. and Peng, Yifan},
  journal={IEEE Transactions on Visualization and Computer Graphics},
  pages={1--13},
  year={2026},
  doi={10.1109/TVCG.2026.3701141}
}

Acknowledgements

The authors thank Dr. Shijie Lin for fruitful discussions. This work was partially supported by the National Science Foundation of China, the Research Grants Council of Hong Kong, and the Innovation and Technology Fund of Hong Kong.

EventTracer: Fast Path Tracing-basedEvent Stream Rendering

EventTracer overview