TeamGL at ACRV Robotic Vision Challenge 1: Probabilistic Object Detection via Staged Non-Suppression Ensembling

This paper describes a novel approach to probabilistic object detection using ensemble techniques. The approach synthesises results from multiple non-probablistic object detectors to acquire final detections. We achieve this by a two-staged ensembling pipeline:(i) identifying detections that are of the same object based on the Intersection over Union (IoU) and labels utilising a greedy assignment process;(ii) creating an ensemble of the detections using a non-suppression algorithm. We employ fixed proportional and label confidence based covariances to capture the spatial uncertainty with particular calibrations on edging objects, a special yet common class of detections. The proposed approach achieved 3rd place in the leaderboard of CVPR-2019 ACRV Robotic Vision Challenge on Probablistic Object Detection.