[Paper Review] StRDAN: Synthetic-to-Real Domain Adaptation Network for Vehicle Re-Identification

Sangrok Lee, Eunsoo Park+2 more|arXiv (Cornell University)|Apr 25, 2020

Domain Adaptation and Few-Shot LearningComputer Science33 references40 citations

TL;DR

This paper proposes StRDAN, a synthetic-to-real domain adaptation network for vehicle re-identification that combines adversarial domain adaptation and semi-supervised learning to leverage large-scale synthetic data with limited real data. By training on both domains using joint and disjoint losses for identity, color, type, and orientation, StRDAN achieves 12.9% and 3.1% mAP improvements over real-data-only baselines on CityFlow-ReID and VeRi datasets, respectively.

ABSTRACT

Vehicle re-identification aims to obtain the same vehicles from vehicle images. This is challenging but essential for analyzing and predicting traffic flow in the city. Although deep learning methods have achieved enormous progress for this task, their large data requirement is a critical shortcoming. Therefore, we propose a synthetic-to-real domain adaptation network (StRDAN) framework, which can be trained with inexpensive large-scale synthetic and real data to improve performance. The StRDAN training method combines domain adaptation and semi-supervised learning methods and their associated losses. StRDAN offers significant improvement over the baseline model, which can only be trained using real data, for VeRi and CityFlow-ReID datasets, achieving 3.1% and 12.9% improved mean average precision, respectively.

Motivation & Objective

• Address the challenge of limited real-world vehicle Re-ID data by leveraging large-scale, inexpensive synthetic data.
• Overcome domain shift between synthetic and real images through adversarial domain adaptation.
• Improve feature learning by exploiting auxiliary labels (color, type, orientation) only available in synthetic data using semi-supervised learning.
• Achieve state-of-the-art performance on benchmark datasets with minimal real data annotation.

Proposed method

• Uses a ResNet-50 backbone for feature extraction from both real and synthetic images.
• Employs a domain discrimination head with adversarial loss to align feature distributions across synthetic and real domains.
• Applies disjoint cross-entropy losses for color, type, and orientation only on synthetic data, enabling semi-supervised learning.
• Combines vehicle ID classification, triplet loss, and domain adversarial loss for end-to-end training.
• Uses a multi-task learning setup with five softmax heads for ID, domain, color, type, and orientation classification.
• Applies data augmentation (horizontal flip, random erasure) and re-ranking post-processing to improve generalization.

Experimental results

Research questions

• RQ1 Can adversarial domain adaptation effectively reduce domain shift between synthetic and real vehicle images in Re-ID?
• RQ2 To what extent does semi-supervised learning using synthetic-only labels (color, type, orientation) improve model generalization on real data?
• RQ3 How do different combinations of disjoint losses (color, type, orientation) affect final Re-ID performance?
• RQ4 Does training on synthetic data with auxiliary labels lead to better feature learning than training only on real data?

Key findings

• StRDAN achieved a 12.9% mAP improvement over the real-data-only baseline on the CityFlow-ReID dataset.
• On the VeRi dataset, StRDAN improved mAP by 3.1% compared to the real-data-only baseline.
• The best performance on CityFlow-ReID was achieved with only orientation supervision (Case 4), while including all three auxiliary labels degraded performance.
• The best performance on VeRi was achieved with only orientation supervision (Case 2), indicating that auxiliary labels may not always improve results.
• The model failed to converge when all disjoint losses were used on the AI City dataset, suggesting potential conflict or overfitting.
• Performance was highly dependent on the real-world dataset, indicating that domain shift and data quality significantly affect model behavior.