LKA-ReID: Vehicle Re-Identification with Large Kernel Attention

Xuezhi Xiang, Zhushan Ma, Lei Zhang, Denis Ombati, Himaloy Himu, Xiantong ZhenICASSP 2024

Abstract

Vehicle Re-Identification (Re-ID) presents significant challenges due to high inter-class similarity and viewpoint variations. This paper introduces LKA-ReID, a novel framework that integrates Large Kernel Attention (LKA) and Hybrid Channel Attention (HCA) within a four-branch network architecture. By decomposing large convolutions to capture global context and employing hybrid attention to isolate discriminative local features, LKA-ReID achieves state-of-the-art performance on the VeRi-776 dataset. Notably, the proposed model significantly reduces parameter counts and computational overhead compared to established baselines.

Introduction

Identifying the same vehicle across disparate camera views is a cornerstone of modern surveillance. However, standard Re-ID methods struggle when different vehicles of the same make and model appear nearly identical from a distance. Furthermore, capturing both global structural information and fine-grained local details (e.g., stickers, scratches, or logos) typically requires secondary models for detection or segmentation, which increases computational costs and data labeling requirements.

LKA-ReID addresses these issues by capturing multi-scale features within a single, efficient end-to-end network, eliminating the need for auxiliary models while maintaining high accuracy.

Methodology

Four-Branch Network Architecture

The core of LKA-ReID is a four-branch architecture designed to process features extracted from a shared backbone. These branches are specialized to handle different levels of feature granularity and attention mechanisms.

The four-branch architecture of LKA-ReID, showing the integration of LKA and HCA modules across different loss branches.

Large Kernel Attention (LKA)

To capture global context without the quadratic complexity of standard self-attention, LKA decomposes a large $K \times K$ convolution into three efficient components:

  1. Depth-wise Convolution: Captures local spatial information.
  2. Depth-wise Dilation Convolution: Captures long-range dependencies to provide global context.
  3. 1×1 Convolution: Refines channel-wise features. This decomposition allows the model to perceive the entire vehicle and its environment efficiently.

Hybrid Channel Attention (HCA)

HCA serves as a discriminative filter, helping the model focus on unique vehicle signatures while ignoring background noise like road surfaces. It combines Global Average Pooling (GAP) to understand overall channel importance with Local Average Pooling (LAP) to preserve critical spatial information within those channels. This dual approach ensures that the model prioritizes regions with the highest identity-defining information.

Experiments and Results

Dataset and Evaluation

The model was evaluated on the VeRi-776 dataset, comprising over 50,000 images of 776 vehicles. Training incorporated metadata such as camera IDs and vehicle viewpoints to enhance robustness.

Performance and Efficiency

LKA-ReID achieved a peak mAP of 86.65% and a Rank-1 accuracy of 98.03%, outperforming recent benchmarks like HCI-Net. A key finding of this research is the efficiency gain provided by the attention modules. As shown in the ablation study, the combination of LKA and HCA not only improves accuracy but also reduces the model's parameter footprint.

Ablation study on the VeRi-776 dataset demonstrating the impact of LKA and HCA on accuracy (mAP, Rank-1) and model efficiency (Params, GFLOPs).

Compared to the baseline, the full LKA-ReID model reduced total parameters from 61.44M to 47.42M (a ~23% reduction) and lowered computational load (GFLOPs) from 14.44 to 13.7.

Conclusion

LKA-ReID demonstrates that sophisticated attention mechanisms can lead to both higher accuracy and greater efficiency in vehicle Re-ID. By leveraging decomposed large kernels and hybrid channel attention, the model effectively distinguishes between visually similar vehicles. Future work will focus on further optimizing the architecture for real-time deployment on edge devices and testing robustness against low-quality imagery where fine-grained details may be obscured.

Summary updated

BibTeX
@inproceedings{xiang2024lka,
  author = {Xuezhi Xiang and Zhushan Ma and Lei Zhang and Denis Ombati and Himaloy Himu and Xiantong Zhen},
  title = {{LKA-ReID: Vehicle Re-Identification with Large Kernel Attention}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2024},
}
Figure