Scholarlink

In order to share image and video datasets for training and validation of machine learning, personal identifiable information (PII) of natural persons needs to be removed. Street scenes are widely used in computer vision research, and anonymization of street scenes typically addresses making faces/heads and license plates unidentifiable. Many existing anonymization approaches are based on bounding boxes of detections, which is problematic as more content than necessary is concealed. This in particular and issue for fisheye cameras, which are quite widely used in vehicle-based capture. We provide a dataset with additional annotations on existing image datasets in order to train a segmentation model for heads and license plates. The training pipeline includes data augmentation in order to handle images resulting from undistorted fisheye images. We train Mask2Former models on the dataset, report the performance of these baseline models, and provide a service implementation and web demo for testing these models.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Training a Segmentation-Based Visual Anonymization Service for Street Scenes

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Martin Korb,
  • Werner Bailer

摘要

In order to share image and video datasets for training and validation of machine learning, personal identifiable information (PII) of natural persons needs to be removed. Street scenes are widely used in computer vision research, and anonymization of street scenes typically addresses making faces/heads and license plates unidentifiable. Many existing anonymization approaches are based on bounding boxes of detections, which is problematic as more content than necessary is concealed. This in particular and issue for fisheye cameras, which are quite widely used in vehicle-based capture. We provide a dataset with additional annotations on existing image datasets in order to train a segmentation model for heads and license plates. The training pipeline includes data augmentation in order to handle images resulting from undistorted fisheye images. We train Mask2Former models on the dataset, report the performance of these baseline models, and provide a service implementation and web demo for testing these models.