Domain Adaptation Frameworks for Road Scene Segmentation in Unseen Environments

未知环境中道路场景分割的领域适应框架

• 批准号: 22K17976
• 负责人: KUMAWAT SUDHAKAR
• 金额: $ 2.75万
• 依托单位: Osaka University
• 依托单位国家: 日本
• 项目类别: Grant-in-Aid for Early-Career Scientists
• 财政年份: 2022
• 资助国家: 日本
• 起止时间: 2022-04-01 至 2024-03-31
• 项目状态: 已结题
• 来源: https://kaken.nii.ac.jp/en/grant/KAKENHI-PROJECT-22K17976/
• 关键词: Domain Generalization; Semantic Segmenattion; Soure Domain; Target Domain; MixStyle; Domain Adptation; Semantic Segmentation

The objective of this research is to develop domain generalization frameworks for the task of semantic segmentation of road scene images using deep learning. We surveyed multiple state-of-the-art existing methods and identified underline hyper-parameters that led to the overfitting of models on the source domains which led to poor generalization. We also identified various techniques that are very useful for domain generalization like mixing styles in feature space and the good range for color jittering paramters like contrast, brightness, saturation, and hue. We observed that our findings are consistent across all the benchmark target domains that we tested. Based on these findings, we have developed an ensemble model for domain generalization which achieves very good performance.

本研究的目的是利用深度学习为道路场景图像的语义分割任务开发领域泛化框架。我们调查了多种最先进的现有方法，并确定了导致源域模型过拟合的下划线超参数，从而导致较差的泛化。我们还确定了各种对领域泛化非常有用的技术，如特征空间中的混合样式，以及对比度、亮度、饱和度和色调等颜色抖动参数的良好范围。我们观察到，我们的发现在我们测试的所有基准目标领域都是一致的。基于这些发现，我们开发了一个集成模型用于领域泛化，并取得了很好的效果。