An Uncertainty Estimation Framework for Probabilistic Object Detection

Zongyao Lyu,Nolan B. Gutierrez,William J. Beksi

DOI: https://doi.org/10.48550/arXiv.2106.15007

2021-06-29

Abstract:In this paper, we introduce a new technique that combines two popular methods to estimate uncertainty in object detection. Quantifying uncertainty is critical in real-world robotic applications. 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 serious repercussions. Our framework employs deep ensembles and Monte Carlo dropout for approximating predictive uncertainty, and it improves upon the uncertainty estimation quality of the baseline method. The proposed approach is evaluated on publicly available synthetic image datasets captured from sequences of video.

Computer Vision and Pattern Recognition

What problem does this paper attempt to address?

The problem that this paper attempts to solve is to quantify uncertainty in object detection. Traditional object detection models may still have ambiguity when providing high - probability outputs, that is, even if the model is very confident in a certain prediction, this prediction may be unreliable. This uncertainty is particularly important in practical applications. For example, the vision system in robotics needs to cope with various operating conditions, such as changes in lighting and occlusions in the surrounding environment. Therefore, the paper proposes a new uncertainty estimation framework, which combines two methods, deep ensembles and Monte Carlo dropout, to approximate the prediction uncertainty, thereby improving the quality of uncertainty estimation of the baseline method. This framework was evaluated on publicly available synthetic image datasets, which were captured from video sequences.

Specifically, the main contributions of the paper include:
- Designing a new aggregation method that combines two popular uncertainty estimation techniques to quantify the uncertainty of object detection models.
- Proposing a new heuristic method to improve the quality of false - positive labels.
- Providing an easy - to - implement method for merging detection ensembles.
- Introducing a high - speed parameter - tuning method to improve the PDQ score.
- Demonstrating that Non - Maximum Suppression (NMS) can be used as a post - processing step to retain all detection results.

Through these methods, the paper aims to improve the reliability and robustness of object detection systems, especially in dynamic scenes and open - set conditions.