Aleatoric Uncertainty Based Label Distribuion Leanig for Pose Estimation of Sichuan Peppers

Sichuan peppers are delicious spices and its autonomous robotic harvesting requires the estimation of pepper pose on the farm. The pepper pose estimation task can be casted as ordinal classification problems by defining a number of pose labels. Since adjacent pose labels share resembling appearance, there exists label ambiguity in many pepper images and ambiguity degree of different images is changing. Label distribution learning (LDL) can handle label ambiguity by converting a single label to a label distribution. Recent LDL methods measure ambiguity degree through training losses, which mainly indicate epistemic uncertainty (EU) and consequently prone to underestimate the ambiguity variations among different samples. To address this issue, we propose a novel LDL method for pepper pose estimation, by measuring ambiguity degrees through aleatoric uncertainty (AU) rather than EU. In specific, we fist leverage Bayesian neural network to quantify sample AU values, which express the inherent noise in observations and thus can sufficiently capture the variations in pose ambiguity. Then, AU values are used to adjust the variance of label distributions so that samples with higher AU values are associated with more flattened distributions. Finally, the instance-aware label distributions are leveraged to train networks for pepper pose estimation. Experimental results on pepper images collected on the farm show that, our method achieve the overall accuracy of 81.28%, outperforming recent leading LDL methods in performance.