Impact of Evidence Theory Uncertainty on Training Object Detection Models

TL;DR

This work addresses training efficiency and robustness for object detection by incorporating Evidence Theory to quantify and leverage prediction uncertainty. It introduces a DS-based fusion of predictions and ground truth to produce an uncertainty measure, which is then mapped to a loss-weighting factor via scorecards and injected into the training loop through DIU/AIU/PIU/Deep strategies. Across binary and multi-class VOC 2012 experiments, the approach consistently improves training efficiency and often improves detection performance, with results analyzed via mAP, confusion matrices, and per-class F1 metrics. The findings demonstrate the practical viability of uncertainty-driven training to accelerate convergence and enhance reliability in object detection, with potential extensions to larger datasets, real-time systems, and multimodal fusion scenarios.

Abstract

This paper investigates the use of Evidence Theory to enhance the training efficiency of object detection models by incorporating uncertainty into the feedback loop. In each training iteration, during the validation phase, Evidence Theory is applied to establish a relationship between ground truth labels and predictions. The Dempster-Shafer rule of combination is used to quantify uncertainty based on the evidence from these predictions. This uncertainty measure is then utilized to weight the feedback loss for the subsequent iteration, allowing the model to adjust its learning dynamically. By experimenting with various uncertainty-weighting strategies, this study aims to determine the most effective method for optimizing feedback to accelerate the training process. The results demonstrate that using uncertainty-based feedback not only reduces training time but can also enhance model performance compared to traditional approaches. This research offers insights into the role of uncertainty in improving machine learning workflows, particularly in object detection, and suggests broader applications for uncertainty-driven training across other AI disciplines.

Figures (19)

• Figure 1: Training with Evidence-Enhanced Loss Function Framework
• Figure 2: Hierarchical Structure of Uncertainty Injection Methods
• Figure 3: Framework of Direct Injection of Uncertainty
• Figure 4: Framework of Average Injection of Uncertainty
• Figure 5: Product Injection of Uncertainty