Sensor Generalization for Adaptive Sensing in Event-based Object Detection via Joint Distribution Training

TL;DR

This paper investigates how intrinsic parameters of bio-inspired event cameras influence object detection performance and proposes a method to improve sensor-agnostic robustness through joint distribution training.

Contribution

It introduces a joint distribution training approach to enhance sensor generalization in event-based object detection models.

Findings

Intrinsic parameters significantly affect detection performance.

Joint distribution training improves sensor-agnostic robustness.

Enhanced model adaptability across different event camera sensors.

Abstract

Bio-inspired event cameras have recently attracted significant research due to their asynchronous and low-latency capabilities. These features provide a high dynamic range and significantly reduce motion blur. However, because of the novelty in the nature of their output signals, there is a gap in the variability of available data and a lack of extensive analysis of the parameters characterizing their signals. This paper addresses these issues by providing readers with an in-depth understanding of how intrinsic parameters affect the performance of a model trained on event data, specifically for object detection. We also use our findings to expand the capabilities of the downstream model towards sensor-agnostic robustness.