Leveraging Temporal Data for Automatic Labelling of Static Vehicles

The development of advanced 3D object detection algorithms for autonomous driving requires a variety of environments to be captured in labelled datasets. While a number of such datasets exist, new ones will continue to be needed to adapt to new domains, sensors and conditions. In this paper, we propose a method to ease the workload of annotators by automatically proposing high-recall labels for static vehicles. We make use of an object detection network pre-trained on an existing dataset to propose detections within a sequence. By determining the location of each frame in a common reference frame, all detections of a static vehicle will share the same location. By averaging these overlapping detections and extending the prediction to all reasonable frames, we generate identical labels for the same object throughout the sequence. We show how our method sequentially refines predictions in order to improve static object recall by over 20% and precision by 7% over an initial set of network proposals.