Event cameras see motion.
- Post by: admin
- June 28, 2026
- No Comment
Event cameras see motion.
Our labels often do not.
Most automotive event-camera datasets still rely on bounding boxes inherited from frame-based pipelines. These annotations tell us where objects are, but not whether they are static or independently moving.
This creates a mismatch between the sensor and the supervision signal.
In our paper “GeoIMO: Geometry-Driven Independent Motion Classification for Event Cameras,” accepted at ICANN 2026, we address this annotation gap with a geometry-driven framework for motion classification.
GeoIMO estimates the ego-motion-induced background motion directly from the event stream using an FOE-based model with yaw compensation. Then, it compares this expected motion with the local motion observed inside each bounding box.
If the object does not follow the expected ego-motion pattern, it is classified as independently moving.
No learning.
No manual motion labels.
Works with any input bounding boxes.
Preprint: https://arxiv.org/abs/2606.24499
Code: https://github.com/smilies-polito/GeoIMO
A special thanks to my co-authors Filippo Marostica, Alessio Caviglia, Alessandro Savino, and Stefano Di Carlo, and to the SMILIES Polito Research Group at Politecnico di Torino for the discussions, feedback, and support throughout this work.
The GIF below shows the idea visually.
hashtagEventBasedVision hashtagComputerVision hashtagAutonomousDriving hashtagNeuromorphicVision hashtagICANN2026
Shared by Anil Bayram Gogebakan
View original post on LinkedIn