Probabilistic Object Detection via Deep Ensembles

摘要

Probabilistic object detection is the task of detecting objects in images and accurately quantifying the spatial and semantic uncertainties of the detection. Measuring uncertainty is important in robotic applications where actions based on erroneous, but high confidence visual detections, can lead to catastrophic consequences. We introduce an approach that employs deep ensembles for estimating predictive uncertainty. The proposed framework achieved 4th place in the ECCV 2020 ACRV Robotic Vision Challenge on Probabilistic Object Detection.