Cross-domain object detection using unsupervised image translation

2021-12-01

Expert systems with applications

AI Summary

This work proposes a concise and interpretable unsupervised domain adaptation method to address the poor generalization of object detection models on unlabeled target domains. By integrating CycleGAN with AdaIN-based image translation, the approach leverages labeled source-domain images and unlabeled target-domain images to generate realistic synthetic target-domain data for training a more robust detector. The generated data effectively bridge the domain gap, substantially narrowing the performance gap compared to models trained with real annotated target-domain data. Extensive experiments demonstrate that the proposed method achieves state-of-the-art performance across multiple autonomous driving benchmarks, highlighting its effectiveness in enhancing cross-domain detection accuracy without requiring target-domain annotations.

Technology Category

• Machine Learning: Transfer, Domain Adaptation, Multi-Task Learning
• Computer Vision: Generative Adversarial Networks (GANs) for Vision
• Application Domains: Mobility, Driving & Flight

Application Category

• Search and Retrieval-Augmented AI: Vertical and domain-specific search
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Economics, Online Markets and Human Computation: Humans versus LLMs for data annotation and labeling

Problem

• Unsupervised domain adaptation
• Object detection
• Cross-domain
• Image translation
• Domain gap

Innovation

• Unsupervised Domain Adaptation
• Object Detection
• Image Translation
• CycleGAN
• AdaIN