Calibrated Aleatoric Uncertainty-Based Adaptive Label Distribution Learning for Pose Estimation of Sichuan Peppers

Pose estimation is crucial to guide a visual harvesting robot to detach crops. In this article, pose estimation for Sichuan peppers is formulated as an ordinal classification problem by defining several pose labels. The shared appearance between neighboring poses results in label ambiguity; meanwhile, there exist obvious variations in ambiguity degrees across images. Conventional one-hot label representations neglect the ambiguity, suffering from overfitting problems. Contrastly, label distribution learning (LDL) methods can handle the pose ambiguity by smoothing a single label to a label distribution. Recent adaptive LDL (ALDL) attempts to construct instance-aware label distributions adaptive to changing ambiguity degrees. However, we find that the existing ALDL methods inevitably underestimate the ambiguity variations in pepper poses. In this article, we devise an ambiguity measure relying on aleatoric uncertainty (AU) and subsequently propose a calibrated AU-based ALDL method for pepper pose estimation. specifically, we start from quantifying AU values for training samples using Bayesian neural networks, where the AU expresses the inherent observation noise. Then, the AU values are calibrated to expected risks heuristically learned on validation sets, preventing AU overestimating the ambiguity. Afterward, the risks act as the ambiguity measure to construct instance-aware label distributions for network training. Experiments on real pepper images show that our method sufficiently captures the ambiguity variations in pepper poses and obtains 88% mean accuracy, outperforming current ALDL methods. Additionally, our method provides reliable assessment on the quality of pose predictions. Both the obtained accuracy and prediction quality are of great practical value to nondestructive harvesting.