EventAugment: Learning Augmentation Policies From Asynchronous Event-Based Data

摘要

Data augmentation is an effective way to overcome the over-fitting problem of deep learning models. However, most existing studies on data augmentation work on frame-like data (e.g., images), and few tackles with event-based data. Event-based data are different from frame-like data, rendering the augmentation techniques designed for frame-like data unsuitable for event-based data. This work deals with data augmentation for event-based object detection and classification, which is important for self-driving, and robot manipulation. Specifically, we introduce EventAugment, a new method to augment asynchronous event-based data by automatically learning augmentation policies. We first identify 13 types of operations for augmenting event-based data. Next, we formulate the problem of finding optimal augmentation policies as a hyperparameter optimization problem. To tackle this problem, we propose a random search-based framework. Finally, we evaluate the proposed method on six public datasets including N-Caltech101, N-Cars, ST-MNIST, N-MNIST, DVSGesture and DDD17. Experimental results demonstrate that EventAugment exhibits substantial performance improvements for both deep neural network-based and spiking neural network-based models, with gains of up to approximately 4%. Notably, EventAugment outperform state-of-the-art methods in terms of overall performance.