In the classical supervised learning settings, classifiers are fit with the assumption of balanced label distributions and produce remarkable results on the same. In the real world, however, these assumptions often bend and in turn adversely impact model performance. Identifying bad learners in skewed target distributions is even more challenging. Thus achieving model robustness under these "label shift" settings is an important task in autonomous perception. In this paper, we analyze the impact of label shift on the task of multi-weather classification for autonomous vehicles. We use this information as a prior to better assess pedestrian detection in adverse weather. We model the classification performance as an indicator of robustness under 4 label shift scenarios and study the behavior of multiple classes of models. We propose t-RAIN a similarity mapping technique for synthetic data augmentation using large scale generative models and evaluate the performance on DAWN dataset. This mapping boosts model test accuracy by 2.1, 4.4, 1.9, 2.7 % in no-shift, fog, snow, dust shifts respectively. We present state-of-the-art pedestrian detection results on real and synthetic weather domains with best performing 82.69 AP (snow) and 62.31 AP (fog) respectively.

翻译：在经典监督学习设置中，分类器基于标签分布平衡的假设进行拟合，并在此假设下取得显著效果。然而在真实世界中，这些假设往往不成立，进而对模型性能产生不利影响。在偏斜的目标分布中识别低性能学习器更具挑战性。因此，在这些"标签偏移"环境下实现模型稳健性成为自主感知领域的重要任务。本文分析了标签偏移对自动驾驶车辆多天气分类任务的影响，并将此信息作为先验知识，以更好地评估恶劣天气下的行人检测性能。我们以分类性能作为指标，研究四种标签偏移场景下的模型稳健性，并探究多类模型的行为特征。提出t-RAIN方法——一种利用大规模生成模型进行合成数据增强的相似度映射技术，并在DAWN数据集上评估其性能。该映射在无偏移、雾、雪、尘暴偏移场景下分别使模型测试准确率提升2.1%、4.4%、1.9%、2.7%。我们给出了真实与合成天气域上最优的行人检测结果，在雪天和雾天场景中分别达到82.69 AP和62.31 AP的最优性能。