Related Experiment Video
Updated: Mar 17, 2026
11:52
Temporal Ordering of Dynamic Expression Data from Detailed Spatial Expression Maps
Published on: February 9, 2017
6.6K
HOTS: A Hierarchy of Event-Based Time-Surfaces for Pattern Recognition.
Xavier Lagorce1, Garrick Orchard2, Francesco Galluppi1
1Vision and Natural Computation Group, Institut National de la Santé et de la Recherche Médicale, Sorbonne Universités, Institut de la Vision, Université Paris 06, Paris, Paris, FranceFrance.
Summary
This study introduces time-surfaces, novel event-based spatio-temporal features for pattern recognition. This biologically inspired hierarchical model achieves high accuracy in character and card recognition tasks.
Area of Science:
- Computer Vision
- Machine Learning
- Computational Neuroscience
Background:
- Existing hierarchical pattern recognition models often rely on traditional frame-based image processing.
- Event-driven vision sensors offer asynchronous, high-temporal-resolution data, mimicking biological systems.
- Extracting meaningful spatio-temporal features from event data remains a challenge.
Purpose of the Study:
- To introduce novel event-based spatio-temporal features called time-surfaces.
- To develop a hierarchical pattern recognition architecture utilizing these time-surfaces.
- To evaluate the model's performance on diverse recognition tasks.
Main Methods:
- Utilized biologically inspired, event-driven vision sensors for data acquisition.
- Developed a hierarchical architecture where successive layers extract increasingly abstract features.
- Employed time-surfaces, representing local temporal activity, as the core feature extraction mechanism.
- Tested the model on character recognition, dynamic card recognition, and moving face recognition tasks.
Main Results:
- Achieved near 100% accuracy on a 36-class character recognition task.
- Achieved near 100% accuracy on a 4-class canonical dynamic card recognition task.
- Obtained 79% accuracy on a new 7-class moving face recognition task.
Conclusions:
- Time-surfaces are robust features for hierarchical event-based pattern recognition.
- The proposed model effectively leverages temporal information from event data.
- The approach demonstrates strong performance across various dynamic visual recognition scenarios.
