针对现有单人姿态估计网络结果缺乏可靠性评估和鲁棒性保障等问题, 提出了一种基于偶然不确定性的测试时增强方法。该方法首先利用随机并行的数据增强和模型推理得到多样化输出, 随后通过计算该输出的偶然不确定性得到其可靠性评估, 最后根据可靠性将该输出及其不确定性进行加权融合以得到更准确鲁棒的结果及其评估。在MPII数据集上的实验表明, 该算法可即插即用地应用于任意现有单人姿态估计网络, 从而得到更精确鲁棒的结果及其不确定性评估。

Aiming at the problem of lacking reliability evaluation and robustness guarantee in existing single-person pose estimation networks′ results, a testing-time-augmentation (TTA) algorithm based on aleatoric uncertainty was proposed. In this TTA algorithm, diverse outputs were firstly obtained by stochastic parallel data augmentation and model inference. Then, the reliability evaluations of those outputs are acquired by calculating their aleatoric uncertainty. Finally, those outputs and their uncertainty were fused according to the reliabilities to obtain a more accurate and robust result as well as its evaluation. Experiments on the MPII dataset show that this algorithm can be applied to any existing single-person pose estimation network in a plug-and-play manner, leading to a more precise and robust result with its uncertainty evaluation.