Highway Driving

高速公路驾驶数据库是视频及其注释的集合，用于语义视频分割。该数据库是在驾驶场景下收集的，对可靠性和实时性要求很高。总共由20个60帧序列组成，帧速率为30Hz。每个视频片段都是用车辆黑匣子的固定位置捕获的。对于每个序列，我们提供手动标记的注释序列，即每个提供的帧都被密集注释。高速公路驾驶数据库提供以下贡献。首先，我们在驾驶场景下提供逼真的视频剪辑。因此，自动驾驶是一个直接受益于我们的数据库结果的应用程序。其次，它提供超过1,000帧注释对。它们被多个注释者反复注释和检查。最后，序列中的注释反映了相邻帧之间的相关性。按时间顺序对单个序列中的帧进行注释，并将先前帧的注释结果提供给注释器作为其参考。由于缺少密集标记的注释，研究语义视频分割任务很麻烦。我们希望高速公路驾驶数据库的时间密集注释将促进有关语义视频分割的有趣的未来研究。

The Highway Driving Database is a collection of videos and their annotations for semantic video segmentation. This database was collected in driving scenarios with high requirements for reliability and real-time performance. It consists of a total of 20 sequences, each containing 60 frames, with a frame rate of 30 Hz. Each video clip is captured from a fixed position on the vehicle's black box. For each sequence, we provide manually labeled annotation sequences, where every supplied frame is densely annotated.

The Highway Driving Database offers the following contributions. First, we provide realistic video clips in driving scenarios, making autonomous driving an application that directly benefits from the outcomes of our database. Second, it provides over 1,000 annotated frame pairs, which have been repeatedly annotated and verified by multiple annotators. Finally, the annotations in the sequences reflect the correlation between adjacent frames. Frames within a single sequence are annotated in chronological order, and the annotation results of previous frames are provided to annotators as references.

Research on the semantic video segmentation task is cumbersome due to the lack of densely labeled annotations. We hope that the temporally dense annotations of the Highway Driving Database will promote interesting future research on semantic video segmentation.

搜集汇总

背景与挑战

背景概述

Highway Driving是一个由OpenDataLab提供的自动驾驶领域数据集，大小为670.2MB，受到较多关注（1.3k）和点赞（142）。