Data Subset selection for training learning models for a variety of tasks, has been widely studied in the literature of batch mode active learning. Recent works attempt to utilize the model specific signals in the deep learning context for computer vision tasks. Companies, in their bid to create safe autonomous driving models, train and test their models on billions of miles of driving data; not all of which may be valuable for a training task. In this paper, we study the problem of frame-subset selection from autonomous vehicle driving data, for the problem of semantic segmentation - which is a crucial component of the perception module in an autonomous driving system. We find that state of the art methods for deep active learning do not utilize pairwise similarity between incoming and existing frames. We explore both active learning settings, where labels for incoming points are not available, as well as frame selection settings and find that our method selects more valuable frames than only score-based frame subset selection, or frame subset selection without label information. We demonstrate the effectiveness of our method using DeeplabV3+ model on both benchmark as well as datasets generated by driving simulators. Our generated dataset and code will be made publicly available.

在批处理模式主动学习的文献中已经广泛研究了用于训练各种任务的学习模型的数据子集选择。最近的工作试图在深度学习上下文中将模型特定的信号用于计算机视觉任务。为了创建安全的自动驾驶模型，公司对数十亿英里的驾驶数据进行了训练和测试；并非所有这些对于培训任务都可能有价值。在本文中，我们研究了从自动驾驶数据中选择框架子集的问题，即语义分割的问题，语义分割是自动驾驶系统感知模块的关键组成部分。我们发现，深度主动学习的最新技术方法没有利用传入帧和现有帧之间的成对相似性。我们探索两种主动学习设置，此处没有用于输入点的标签以及帧选择设置，并且发现我们的方法选择的帧比仅基于得分的帧子集选择或没有标签信息的帧子集选择更多。我们在基准测试以及驾驶模拟器生成的数据集上展示了使用DeeplabV3 +模型的方法的有效性。我们生成的数据集和代码将公开提供。