视频异常分割数据集

本数据集是专为自动驾驶设计的第一个视频异常分割数据集，由清华大学等机构创建。数据集包含120,000个高分辨率帧，分布在7个不同的城镇，涵盖城市和农村地区，使用CARLA模拟器生成。数据集包含200个视频序列，每个序列时长10秒，帧率为60 FPS，每帧分辨率为1920×1080，并配有精确的像素级异常、语义和实例地图。此外，数据集还记录了渲染G缓冲区，用于增强合成数据的真实感。数据集旨在解决自动驾驶中异常对象的及时准确分割问题，提高驾驶安全性，特别是在需要人类干预的情况下。

This is the first video anomaly segmentation dataset tailored for autonomous driving, co-created by Tsinghua University and other institutions. The dataset contains 120,000 high-resolution frames, generated using the CARLA simulator across 7 distinct towns spanning both urban and rural areas. It includes 200 video sequences, each with a duration of 10 seconds, a frame rate of 60 FPS, and a resolution of 1920×1080 per frame, paired with precise pixel-level anomaly, semantic, and instance maps. Additionally, the dataset provides recorded rendered G-buffers to enhance the realism of the synthetic data. This dataset aims to address the challenge of timely and accurate segmentation of anomalous objects in autonomous driving, thereby improving driving safety, particularly in scenarios requiring human intervention.