This paper investigates spurious features in a purpose-built image dataset of Hemudu pottery sherds. To minimize acquisition bias, we constructed a standardized dataset by photographing the fragments in situ at the excavation site, capturing diverse cues such as texture, edges, and shape. Our experiments reveal that fine-grained shape details, though correlated with class labels, function as spurious predictors and undermine robustness. To address this, we propose a supervised contrastive learning method, distinct from deep feature reweighting (DFR). By aligning each sample with its label-preserving augmentations, in which spurious cues are disrupted, we promote augmentation-invariant representations. We further adapt the contrastive loss to more effectively balance positive and negative pairs, and incorporate the flooding strategy to stabilize training. On our Hemudu sherd dataset, the method achieves 97.3% accuracy, outperforming recent baselines and demonstrating its effectiveness in mitigating reliance on spurious shortcuts.
Yu et al. (Wed,) studied this question.