Uncertainty aware stochastic sampling for efficient object detection

The effectiveness of deep learning models is strongly influenced by the quality of training data. Traditional training approaches assume that all samples contribute equally to the learning process, leading to uniform data sampling. However, this assumption overlooks the substantial variation in informational content across samples. This paper introduces a novel and computationally lightweight data prioritization methodology for object detection that dynamically adjusts the sampling probability of training data according to its relevance during learning.The proposed Relative Detection Error (RDE) is a new temporal instability metric defined in a joint classification-regression setting, where classification corresponds to object category prediction and regression corresponds to bounding-box parameter estimation. By explicitly quantifying prediction variability over time, RDE identifies samples with higher learning value and guides an exploration-regularized stochastic sampling policy. As a result, the method improves both classification and localization accuracy while imposing minimal computational overhead and integrating seamlessly into standard training pipelines. The approach is validated using YOLO architectures on diverse datasets, demonstrating consistent improvements and strong generalization across models and domains. Experimental results show that prioritizing high-value samples yields higher F1 scores and mean Average Precision, alongside more efficient and stable convergence. The project source code is available at: https://kp-labs-bme.github.io/Object-Detection-Prioritization/.