[Paper Review] EventDrop: data augmentation for event-based learning

Fuqiang Gu, Weicong Sng+2 more|arXiv (Cornell University)|Jun 7, 2021

Advanced Memory and Neural Computing28 references4 citations

TL;DR

This paper proposes EventDrop, a simple and computationally efficient data augmentation method for event-based learning that improves model generalization by dropping events using three strategies: random, time-based, and area-based dropping. It significantly boosts classification accuracy across multiple deep networks on N-Caltech101 and N-Cars datasets, with up to 95.5% accuracy on EST representation using ResNet-34.

ABSTRACT

The advantages of event-sensing over conventional sensors (e.g., higher dynamic range, lower time latency, and lower power consumption) have spurred research into machine learning for event data. Unsurprisingly, deep learning has emerged as a competitive methodology for learning with event sensors; in typical setups, discrete and asynchronous events are first converted into frame-like tensors on which standard deep networks can be applied. However, over-fitting remains a challenge, particularly since event datasets remain small relative to conventional datasets (e.g., ImageNet). In this paper, we introduce EventDrop, a new method for augmenting asynchronous event data to improve the generalization of deep models. By dropping events selected with various strategies, we are able to increase the diversity of training data (e.g., to simulate various levels of occlusion). From a practical perspective, EventDrop is simple to implement and computationally low-cost. Experiments on two event datasets (N-Caltech101 and N-Cars) demonstrate that EventDrop can significantly improve the generalization performance across a variety of deep networks.

Motivation & Objective

• To address overfitting in event-based deep learning due to small dataset sizes.
• To improve model generalization by increasing training data diversity without increasing data collection costs.
• To develop a lightweight, input-space data augmentation method specifically tailored for asynchronous event data.
• To evaluate the effectiveness of event dropping strategies across different event representations and network architectures.

Proposed method

• EventDrop introduces three event dropping strategies: Random drop, Drop by time, and Drop by area, each designed to simulate occlusion or sensor noise.
• Events are dropped during training with a fixed probability (0.5) per strategy, and the magnitude of dropping is randomly sampled from a predefined set.
• The method operates directly on input event data before feeding into deep networks, making it compatible with both DNNs and SNNs.
• EventDrop combines all three strategies with equal probability, enabling dynamic augmentation per training sample.
• The approach is computationally efficient and requires no additional learnable parameters, making it easy to integrate into existing training pipelines.
• Different event representations—event frame, event count, voxel grid, and EST—are used to evaluate the method’s robustness across common input formats.

Experimental results

Research questions

• RQ1 Can dropping events in a structured way improve generalization in event-based deep learning?
• RQ2 How do different event dropping strategies (random, time-based, area-based) affect model performance across various event representations?
• RQ3 Does EventDrop outperform standard baselines and existing augmentation techniques in terms of accuracy and robustness on real-world and simulated event datasets?
• RQ4 How does the amount of training data influence the effectiveness of EventDrop compared to baseline methods?

Key findings

• EventDrop achieves a test accuracy of 95.50% on the N-Cars dataset using the EST representation with ResNet-34, significantly outperforming the baseline of 91.03%.
• On the N-Caltech101 dataset, EventDrop achieves 85.15% accuracy with EST representation, compared to 83.91% for the baseline.
• The method consistently improves performance across all event representations and network architectures tested, including ResNet-34 and VGG-16.
• Even with only 10% of the training data, EventDrop achieves ~94% accuracy on N-Cars, compared to ~91.5% for the baseline, demonstrating robustness to data scarcity.
• Drop by area and EventDrop show the strongest improvements on the real-world N-Cars dataset, suggesting effectiveness in handling sensor noise and occlusion.
• The performance gain is most pronounced when using the EST representation, indicating that EventDrop is particularly effective with high-fidelity event representations.