E2Detect: Object Detection from Event Camera via Sparse Feature Pyramid Recovery
Year: 2026
Type: conference-paper
Abstract: Event cameras offer high dynamic range, no motion blur, and very high temporal resolution—capabilities that conventional RGB cameras lack and can be crucial for many vision tasks. Existing computer vision approaches struggle to operate directly on event streams due to the asynchronous and sparse nature of event data. Event-driven pipelines often require intensity image reconstruction or retraining on event-based object detection datasets, increasing the computational complexity and limiting the reusability of existing pretrained frame-based object detectors. We present E2Detect, a learning-based framework that performs object detection directly from event streams without dense intensity image reconstruction or detector retraining. Our Event Feature Pyramid Network (E-FPN) recovers a Single Shot MultiBox Detector (SSD)-compatible feature pyramid directly from event data. The recovered feature pyramid is plug-and-play compatible with the SSD detection head without requiring retraining. Experiments on ESIM-simulated event data from the REDS 120fps dataset demonstrate that our method achieves competitive [email protected] as compared with RGB image–based detection results.
Authors: Chris Henry, Zhu Li, Aggelos K. Katsaggelos
Institutions: University of Missouri–Kansas City, Northwestern University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7165194805