The paper introduces Detecting Every Object in Events (DEOE), a novel approach for class-agnostic open-world object detection using event cameras. Event cameras offer several advantages over traditional frame-based sensors, including sub-millisecond latency and high dynamic range, making them well-suited for robust object detection in challenging scenarios.
The key components of DEOE are:
Spatio-temporal Consistency: The Dual Regressor Head measures the spatial and temporal consistency of object proposals to identify potential unknown objects. Samples with high spatial and temporal IoU are considered potential positive samples and are incorporated into the training process.
Task Disentanglement: The Disentangled Objectness Head decouples the foreground-background classification and unknown object discovery tasks. One branch focuses on distinguishing annotated foreground from background, while the other branch learns to identify potential foreground objects, including unknown categories.
The authors conduct extensive experiments on the 1 Megapixel Automotive Detection Dataset, demonstrating the superiority of DEOE over several strong baselines that integrate state-of-the-art event-based object detectors with advancements in RGB-based class-agnostic object detection. DEOE achieves high detection performance on both known and unknown classes while maintaining a rapid detection speed.
The paper also includes a cross-dataset evaluation on the DSEC-Detection dataset, further validating DEOE's generalization capabilities. The qualitative results showcase DEOE's ability to detect objects missed by RGB-based methods, particularly in extreme scenarios involving fast-moving objects and challenging lighting conditions.
To Another Language
from source content
arxiv.org
Key Insights Distilled From
by Haitian Zhan... at arxiv.org 04-09-2024
https://arxiv.org/pdf/2404.05285.pdfDeeper Inquiries