InstanceMotSeg：用于自动驾驶的实时实例运动分割

运动对象分割是自动驾驶汽车的一项关键任务，因为它可用于基于对象的运动线索以不可知的方式分割对象。它可以根据训练的运动来检测未知物体（例如，驼鹿或工程车）。..

InstanceMotSeg: Real-time Instance Motion Segmentation for Autonomous Driving

Moving object segmentation is a crucial task for autonomous vehicles as it can be used to segment objects in a class agnostic manner based on their motion cues. It enables the detection of unknown objects during training (e.g., moose or a construction truck) based on their motion.Although pixel-wise motion segmentation has been studied in autonomous driving literature, it is not dealt with on the instance level, which would help separate connected segments of moving objects leading to better trajectory planning. Other generic video object segmentation tasks have dealt with instance-wise motion segmentation but on much simpler setting and have not dealt with the multi-task learning problem for both semantic and class agnostic instance segmentation. In this paper, we propose a motion-based instance segmentation task and provide a new annotated dataset based on KITTIMoSeg, which will be released publicly. Our dataset provide extra class annotations which is crucial for studying class agnostic segmentation. We further propose an efficient multi-task learning approach that learns an extra class agnostic instance segmentation head through sharing the prototype generation network with the semantic head. The model then learns separate prototype coefficients within the class agnostic and semantic heads. To obtain real-time performance, we study different efficient encoders and obtain 39 fps on a Titan Xp GPU using MobileNetV2 with an improvement of 10% mAP relative to the baseline. Our model outperforms state of the art motion segmentation methods with 3.3% improvement. We summarize our work in a short video with qualitative results at https://youtu.be/3mOeDmiR8BU.