Technology Category
Computer Vision: Adversarial Attacks & RobustnessMachine Learning: Adversarial Learning & RobustnessIntelligent Robots: Multimodal Perception & Sensor Fusion
Learning from Noise: Enhancing DNNs for Event-Based Vision through Controlled Noise Injection
🤖 AI Summary
Event camera data is highly susceptible to noise, which severely degrades the robustness of deep learning models; conventional post-hoc filtering methods often erroneously discard informative events. To address this, we propose a controllable noise-injection training paradigm specifically designed for event data—eliminating reliance on post-processing filters. Our approach explicitly models impulse noise during training, constructs multi-granularity event representations, and introduces a cross-architecture robust training framework. Notably, this is the first work to demonstrate strong generalization robustness across diverse architectures—including CNNs, Vision Transformers (ViTs), Spiking Neural Networks (SNNs), and Graph Convolutional Networks (GCNs). Extensive experiments on N-Caltech101, N-Cars, and Mini N-ImageNet show that our method achieves the highest average classification accuracy and exhibits superior stability under dynamically varying noise intensities, significantly outperforming state-of-the-art event filtering techniques.
Technology Category
Application Category
📝 Abstract
Event-based sensors offer significant advantages over traditional frame-based cameras, especially in scenarios involving rapid motion or challenging lighting conditions. However, event data frequently suffers from considerable noise, negatively impacting the performance and robustness of deep learning models. Traditionally, this problem has been addressed by applying filtering algorithms to the event stream, but this may also remove some of relevant data. In this paper, we propose a novel noise-injection training methodology designed to enhance the neural networks robustness against varying levels of event noise. Our approach introduces controlled noise directly into the training data, enabling models to learn noise-resilient representations. We have conducted extensive evaluations of the proposed method using multiple benchmark datasets (N-Caltech101, N-Cars, and Mini N-ImageNet) and various network architectures, including Convolutional Neural Networks, Vision Transformers, Spiking Neural Networks, and Graph Convolutional Networks. Experimental results show that our noise-injection training strategy achieves stable performance over a range of noise intensities, consistently outperforms event-filtering techniques, and achieves the highest average classification accuracy, making it a viable alternative to traditional event-data filtering methods in an object classification system. Code: https://github.com/vision-agh/DVS_Filtering
Problem
Research questions and friction points this paper is trying to address.
Enhancing DNN robustness against event data noise
Improving noise-resilient representations in event-based vision
Replacing traditional event-filtering with noise-injection training
Innovation
Methods, ideas, or system contributions that make the work stand out.
Noise-injection training enhances DNN robustness
Controlled noise added to training data
Outperforms traditional event-filtering techniques
🔎 Similar Papers
No similar papers found.
