Uncertainty-aware Instance Segmentation using Dropout Sampling

Vision is an integral part of many robotic systems, and especially so when a robot must interact with its environment. In such cases, decisions made based on erroneous visual detections can have disastrous consequences. Hence, being able to accurately measure the uncertainty associated with visual information is highly important for making informed decisions. However, this uncertainty is often not captured by classic computer vision systems or metrics. In this paper we address the task of instance segmentation in a robotics context, where we are concerned with uncertainty associated with not only the class of an object (semantic uncertainty) but also its location (spatial uncertainty). We apply dropout sampling to the state-of-the-art instance segmentation network Mask-RCNN to provide estimates of both semantic uncertainty and spatial uncertainty. We show that a metric which combines both uncertainty measures provides an estimate of uncertainty which improves over either one individually. Additionally, we apply our technique to the ACRV Probabilistic Object Detection dataset where it achieves a score of 14.65.