Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

AW-GATCN: Adaptive Weighted Graph Attention Convolutional Network for Event Camera Data Joint Denoising and Object Recognition

Haiyu Li, Charith Abhayaratne

TL;DR

Event cameras produce sparse, high-temporal-resolution data with substantial noise, complicating denoising and object recognition. The authors introduce AW-GATCN, an Adaptive Weighted Graph Attention Convolutional Network that unifies adaptive event-point segmentation, multi-factor edge weighting, and adaptive graph-denoising within a graph attention framework to robustly model spatiotemporal structures. Key contributions include a four-factor edge weight formulation, density-driven adaptive segmentation and voxelization, a degree-distribution–driven thresholding scheme for denoising, and an attention-based GCN that emphasizes informative connections; these yield state-of-the-art accuracies on four datasets, with notable improvements in noise reduction and a substantial gain over Euclidean-only approaches. Across N-Caltech101, CIFAR10-DVS, MNIST-DVS, and N-CARS, AW-GATCN achieves 83.77%, 76.79%, 99.30%, and 96.89% accuracy, respectively, surpassing existing graph-based methods by up to 8.79% and improving denoising performance by up to 19.57% (and up to 6.26% relative to Euclidean baselines), highlighting its practical impact for robust event-based recognition in noisy, heterogeneous environments.

Abstract

Event cameras, which capture brightness changes with high temporal resolution, inherently generate a significant amount of redundant and noisy data beyond essential object structures. The primary challenge in event-based object recognition lies in effectively removing this noise without losing critical spatial-temporal information. To address this, we propose an Adaptive Graph-based Noisy Data Removal framework for Event-based Object Recognition. Specifically, our approach integrates adaptive event segmentation based on normalized density analysis, a multifactorial edge-weighting mechanism, and adaptive graph-based denoising strategies. These innovations significantly enhance the integration of spatiotemporal information, effectively filtering noise while preserving critical structural features for robust recognition. Experimental evaluations on four challenging datasets demonstrate that our method achieves superior recognition accuracies of 83.77%, 76.79%, 99.30%, and 96.89%, surpassing existing graph-based methods by up to 8.79%, and improving noise reduction performance by up to 19.57%, with an additional accuracy gain of 6.26% compared to traditional Euclidean-based techniques.

AW-GATCN: Adaptive Weighted Graph Attention Convolutional Network for Event Camera Data Joint Denoising and Object Recognition

TL;DR

Event cameras produce sparse, high-temporal-resolution data with substantial noise, complicating denoising and object recognition. The authors introduce AW-GATCN, an Adaptive Weighted Graph Attention Convolutional Network that unifies adaptive event-point segmentation, multi-factor edge weighting, and adaptive graph-denoising within a graph attention framework to robustly model spatiotemporal structures. Key contributions include a four-factor edge weight formulation, density-driven adaptive segmentation and voxelization, a degree-distribution–driven thresholding scheme for denoising, and an attention-based GCN that emphasizes informative connections; these yield state-of-the-art accuracies on four datasets, with notable improvements in noise reduction and a substantial gain over Euclidean-only approaches. Across N-Caltech101, CIFAR10-DVS, MNIST-DVS, and N-CARS, AW-GATCN achieves 83.77%, 76.79%, 99.30%, and 96.89% accuracy, respectively, surpassing existing graph-based methods by up to 8.79% and improving denoising performance by up to 19.57% (and up to 6.26% relative to Euclidean baselines), highlighting its practical impact for robust event-based recognition in noisy, heterogeneous environments.

Abstract

Event cameras, which capture brightness changes with high temporal resolution, inherently generate a significant amount of redundant and noisy data beyond essential object structures. The primary challenge in event-based object recognition lies in effectively removing this noise without losing critical spatial-temporal information. To address this, we propose an Adaptive Graph-based Noisy Data Removal framework for Event-based Object Recognition. Specifically, our approach integrates adaptive event segmentation based on normalized density analysis, a multifactorial edge-weighting mechanism, and adaptive graph-based denoising strategies. These innovations significantly enhance the integration of spatiotemporal information, effectively filtering noise while preserving critical structural features for robust recognition. Experimental evaluations on four challenging datasets demonstrate that our method achieves superior recognition accuracies of 83.77%, 76.79%, 99.30%, and 96.89%, surpassing existing graph-based methods by up to 8.79%, and improving noise reduction performance by up to 19.57%, with an additional accuracy gain of 6.26% compared to traditional Euclidean-based techniques.
Paper Structure (19 sections, 12 equations, 3 figures, 3 tables, 2 algorithms)

This paper contains 19 sections, 12 equations, 3 figures, 3 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. Adaptive Event Point Segmentation
  5. Normalized Density
  6. Determination of Adaptive Window Size
  7. Determination of Adaptive Voxel Count
  8. Adaptive Denoising
  9. Custom Weight Calculation
  10. Dynamic Threshold Adjustment Based on Normalized Degree Distribution
  11. AW-GATCN Network Architecture
  12. Graph Convolutional Layer with Attention Mechanism
  13. Performance Evaluation
  14. Experimental Setup
  15. Comparison with Graph-based Methods for Object Recognition
  16. ...and 4 more sections

Figures (3)

  • Figure 1: The adaptive edge weight $w_{ij}$, designed to capture event point relevance, facilitates noise filtering and attention-based feature aggregation, enhancing robust recognition by emphasizing the most informative node connections.
  • Figure 2: The circular region represents the optimal threshold $T$ determined by maximizing variance. Purple nodes indicate high-correlation points with weights $w_{ij}$ less than $T$, while teal nodes are excluded.
  • Figure 3: Compared to Original Data, Comb 3 effectively reduces the number of event points while retaining the primary structure of the recognized object.