BiasBench: A reproducible benchmark for tuning the biases of event cameras

TL;DR

BiasBench addresses the need for reproducible bias tuning in event cameras by introducing a benchmark dataset with grid-like bias settings across three scenes and scene-specific quality metrics. The authors frame bias tuning as an imitation-learning problem and provide a TD3+BC baseline that learns online bias adjustments from expert demonstrations, using a ResNet50-based feature extractor on accumulated ON/OFF event frames. The dataset enables evaluation of how biases influence downstream tasks (e.g., tracking, frequency estimation, VO) beyond simple event-rate metrics, and demonstrates that bias choice materially affects task performance even when raw event streams look similar. The work facilitates systematic development of bias-tuning algorithms and highlights practical design choices, such as accumulation time and feature extractor, that impact learning efficiency and generalization. Overall, BiasBench offers a practical, reproducible path toward automated bias configuration for real-world event-camera applications, with potential to accelerate robustness and performance in robotics and navigation tasks.

Abstract

Event-based cameras are bio-inspired sensors that detect light changes asynchronously for each pixel. They are increasingly used in fields like computer vision and robotics because of several advantages over traditional frame-based cameras, such as high temporal resolution, low latency, and high dynamic range. As with any camera, the output's quality depends on how well the camera's settings, called biases for event-based cameras, are configured. While frame-based cameras have advanced automatic configuration algorithms, there are very few such tools for tuning these biases. A systematic testing framework would require observing the same scene with different biases, which is tricky since event cameras only generate events when there is movement. Event simulators exist, but since biases heavily depend on the electrical circuit and the pixel design, available simulators are not well suited for bias tuning. To allow reproducibility, we present BiasBench, a novel event dataset containing multiple scenes with settings sampled in a grid-like pattern. We present three different scenes, each with a quality metric of the downstream application. Additionally, we present a novel, RL-based method to facilitate online bias adjustments.

Figures (24)

• Figure 1: Accumulated event frames of the spinning disk setup of our event dataset with different bias_diff_on and bias_diff_off. As can be seen, different bias settings heavily influence the signal-to-noise ratio.
• Figure 2: The spatter tracker from the Metavision SDK for different event data with the bias settings (left) $(20, -10, -35, 24, 126)$, (middle) $(20, 25, 10, 0, -20)$, and (right) $(-10, 95, 10, 0, 53)$.
• Figure 3: Blinking LED setup
• Figure 4: Examples of the frequency estimation for different LEDs and bias settings. The real event data is given in blue, the best fit in red. The red-shaded area resembles the uncertainty of the fit from the frequency estimate. The resulting Relative Frequency Uncertainty is for the function on the top $0.018$, for the function in the middle $0.04$, and for the function on the bottom $0.4$.
• Figure 5: The ground truth trajectory of the end-effector of a Pandas robot arm is shown in in black. The trajectory output from DEVO Klenk20243dv for the bias settings $(17, 18, 33, 24)$ is shown in in blue. And the trajectory output for the bias settings $(17, 18, 55, 96)$ is shown in in green. Despite the similar bias values, the VO failed in one case.