Overview

Object detection is a central task in computer vision. Quantifying detection uncertainty is critical for real-world robotic applications. Robotic vision systems must be able to cope with diverse operating conditions such as variations in illumination and occlusions in the surrounding environment. Traditional detection models can be ambiguous even when they provide a high-probability output. Robot actions based on high-confidence, yet unreliable predictions, may result in catastrophic outcomes. Models for object detection have achieved high mean average precision (mAP) on datasets such as ImageNet, PASCAL Visual Object Classes, and Microsoft Common Objects in Context. However, these models can fail when evaluated in dynamic scenarios that contain objects from outside the training dataset. The mAP metric encourages detectors to output many detections for each image, yet it does not provide a consistent measure of confidence other than a higher score.

Contributions

• We created a deep ensemble architecture that made it to the finals of the Probabilistic Object Detection Challenge
• We developed an object detection framework that employs deep ensembles and Monte Carlo dropout for approximating predictive uncertainty

Publications

1. Z. Lyu, N. Gutierrez, A. Rajguru, and W.J. Beksi, "Probabilistic Object Detection via Deep Ensembles," European Conference on Computer Vision (ECCV) Workshops, 2020.
   Paper  •   Preprint  •   Citation
2. Z. Lyu, N.B. Gutierrez, and W.J. Beksi, "An Uncertainty Estimation Framework for Probabilistic Object Detection," International Conference on Automation Science and Engineering (CASE), 2021.
   Paper  •   Preprint  •   Source Code  •   Citation