[Paper Review] Eliminating cross-camera bias for vehicle re-identification

Jinjia Peng, Guangqi Jiang+4 more|arXiv (Cornell University)|Dec 21, 2019

Video Surveillance and Tracking Methods42 references4 citations

TL;DR

This paper proposes a Cross-Camera Adaptation (CCA) framework and an Attention Alignment Network (AANet) to reduce cross-camera bias in vehicle re-identification by transferring images to a common camera domain and focusing on discriminative vehicle parts. The method achieves state-of-the-art performance, improving mAP by 2.6% and rank-1 accuracy by 1.65% on VeRi-776 compared to baseline models.

ABSTRACT

Vehicle re-identification (reID) often requires recognize a target vehicle in large datasets captured from multi-cameras. It plays an important role in the automatic analysis of the increasing urban surveillance videos, which has become a hot topic in recent years. However, the appearance of vehicle images is easily affected by the environment that various illuminations, different backgrounds and viewpoints, which leads to the large bias between different cameras. To address this problem, this paper proposes a cross-camera adaptation framework (CCA), which smooths the bias by exploiting the common space between cameras for all samples. CCA first transfers images from multi-cameras into one camera to reduce the impact of the illumination and resolution, which generates the samples with the similar distribution. Then, to eliminate the influence of background and focus on the valuable parts, we propose an attention alignment network (AANet) to learn powerful features for vehicle reID. Specially, in AANet, the spatial transfer network with attention module is introduced to locate a series of the most discriminative regions with high-attention weights and suppress the background. Moreover, comprehensive experimental results have demonstrated that our proposed CCA can achieve excellent performances on benchmark datasets VehicleID and VeRi-776.

Motivation & Objective

• Address the significant cross-camera bias in vehicle re-identification caused by variations in illumination, viewpoint, resolution, and background.
• Reduce domain shift between cameras by transferring multi-camera images into a common reference camera space to normalize appearance differences.
• Improve feature learning by focusing on discriminative vehicle parts and suppressing background interference.
• Develop a unified framework that combines domain adaptation and attention-based feature extraction for robust vehicle re-identification.
• Achieve state-of-the-art performance on benchmark datasets VeRi-776 and VehicleID by mitigating camera-specific biases.

Proposed method

• Utilize StarGAN-based image translation to transfer vehicle images from multiple cameras into a single reference camera domain, aligning illumination, resolution, and background.
• Introduce an Attention Alignment Network (AANet) that employs a Spatial Transformer Network (STN) with an attention module to localize and focus on high-discriminative regions of the vehicle.
• Apply the STN with attention to dynamically identify and extract salient features from vehicle images, reducing background interference.
• Train the model using a multi-branch architecture that learns global features, part-based features, and fused features for improved representation.
• Optimize the network using cross-camera contrastive loss and triplet loss to enhance feature discrimination across domains.
• Apply domain adaptation via image translation to reduce distribution shift between cameras, enabling better generalization across diverse camera views.

Experimental results

Research questions

• RQ1 How can cross-camera appearance bias caused by illumination, viewpoint, and resolution differences be effectively reduced in vehicle re-identification?
• RQ2 To what extent does domain adaptation via image translation improve feature generalization across multi-camera surveillance systems?
• RQ3 Can attention-based localization of vehicle parts improve feature discriminability compared to global feature extraction?
• RQ4 How does combining global, part-based, and fused features affect re-identification accuracy?
• RQ5 Does the proposed method outperform existing state-of-the-art approaches on standard benchmarks like VeRi-776 and VehicleID?

Key findings

• The proposed CCA framework improves mAP by 2.6% and rank-1 accuracy by 1.65% on VeRi-776 compared to the original AANet without domain adaptation.
• The AANet with attention-based localization increases mAP by 2.34% and rank-1 by 2.68% on VeRi-776 compared to rigid part-based splitting.
• On VehicleID, the method achieves a 3.61% mAP improvement on test sets of size 3200, demonstrating robustness to varying data scales.
• Fusing global and part-based features yields a 11.34% mAP gain over global features alone on VeRi-776, confirming the benefit of multi-scale feature learning.
• Visualization results show high retrieval accuracy and robustness across diverse viewpoints and lighting conditions.
• The ablation study confirms that both the image translation (domain adaptation) and attention-based localization are essential for performance gains.