Mitigating Contextual Bias

The contextual bias split of the Aircraft dataset was constructed by Dunlap et al. [https://arxiv.org/abs/2305.16289]. The split uses two visually similar classes: Boeing-767 and Airbus-322. Each image in this split is categorized as “sky”, “grass”, or “road” depending on its background, with ambiguous examples filtered out. The bias in the dataset is introduced by training on 400 samples where Airbus aircraft are exclusively associated with road backgrounds and Boeing aircraft with grass backgrounds, although both types may appear against sky backgrounds.