Contrastive pretraining for semantic segmentation is robust to noisy positive pairs

Domain-specific variants of contrastive learning can construct positive pairs from two distinct images, as opposed to augmenting the same image twice. Unlike in traditional contrastive methods, this can result in positive pairs not matching perfectly. Similar to false negative pairs, this could impede model performance. Surprisingly, we find that downstream semantic segmentation is either robust to the noisy pairs or even benefits from them. The experiments are conducted on the remote sensing dataset xBD, and a synthetic segmentation dataset, on which we have full control over the noise parameters. As a result, practitioners should be able to use such domain-specific contrastive methods without having to filter their positive pairs beforehand.