Uncertainty-Aware Rotation Estimation for YOLO-Based Oriented Object Detection

This work proposes a practical, easily integrable method to estimate the measurement uncertainty associated with the rotation angle predicted by an oriented object detector based on YOLO-OBB. The approach is based on Monte Carlo (MC) Dropout during inference, made possible by the inclusion of Dropout2d layers at the input of the detection head, injected before training starts, of the conventional YOLO architecture. This architectural modification is applied in a non-invasive manner, preserving the original training pipeline and allowing adoption in real-world scenarios with negligible per-pass overhead, total runtime scales approximately linearly with the number of stochastic passes. During inference, the model performs multiple stochastic forward passes, producing a set of predictions for each object. To ensure consistency across sampled predictions, the oriented bounding boxes are matched across passes using an IoU-based clustering mechanism, which groups estimates referring to the same target and filters out potential spurious detections. The angular uncertainty is then quantified through statistics of the sampled estimates, the mean and standard deviation of the predicted angles. Experimental results were obtained on a dataset of electronic components on PCBs as an application-relevant case study. Overall, the proposed method provides a lightweight, deployable uncertainty mechanism without changing the loss function or training multiple models (single standard training run), making it suitable for orientation-dependent applications with stringent requirements in terms of reliability and quality control.