报告人： 周千寓

导师： 马利庄 教授

时间： 202 4 年 3 月 1 日（星期 五 ） 10:50-11:30

地点： 电院 3号楼404会议室

【报告题目】

复杂场景下的视觉迁移学习算法研究

【报告摘要】

视觉迁移学习旨在提升现有视觉模型面向未知目标域的迁移性和泛化性，在智能驾驶、人脸安全、城市治理等国家重大需求领域具有广阔的应用前景，并有着重要的学术研究价值。 本报告主要聚焦于 复杂场景下环境多变监督少、时空差异干扰大、数据受限访问难的关键问题，并突破跨域特征对齐难、运动目标建模难、未知场景推断难的技术瓶颈，实现复杂场景下的视觉领域自适应与分布外泛化 。 具体提出了三项成果： 基于场景知识先验的跨域特征对齐、基于时空 Transformer的运动目标建模、基于自适应域特征学习的受限访问迁移方法，适用于语义分割、目标检测、人脸活体检测任务，提升了现有模型的适应性、时效性、鲁棒性。

【 Abstract】

Visual transfer learning aims to improve the transferability and g eneralizability of existing visual models for unknown target domains, which has broad application prospects in major national demand fields such as intelligent driving, face security, and urban governance, and has important academic research value. This report mainly focuses on the key problems of limited supervision due to the varying environments , large interference of spati al -temporal differences, and limited access to data in complex scenes, and breaks through the technical bottlenecks of difficulty in cross-domain feature alignment, moving object modeling, and unknown scene inference, and realizes visual domain adaptation and out-of-distribution generalization in complex scenes. Specifically, we proposes three achievements: cross-domain feature alignment based on scene knowledge prior, moving object modeling based on spati al -temporal Transformer, and restricted access transfer method based on adaptive domain -invariant feature learning, which is applicable for semantic segmentation, object detection, and face -anti spoofing tasks, and improves the adaptability, timeliness, and robustness of existing models.