DaCo: domain-agnostic contrastive learning for visual place recognition

Visual place recognition is a core component of visual information analysis, which serves for the position and orientation perception of autonomous driving and robotics. The current place recognition methods usually rely on image retrieval techniques to identify the visual similarity between query and gallery images. However, state-of-the-art image retrieval methods are often based on extensive labels, such as matched pairs (e.g., the image correspondences). Besides, image retrieval methods heavily suffer from environmental condition changes (i.e., a large range of illumination and weather changes). To alleviate the annotation cost, we introduce contrastive learning to perform feature extraction and feature similarity measurement in a self-supervised manner. Considering the heavy data augmentations of the existing contrastive learning approaches cannot effectively simulate domain disparities, we design the generative adversarial model to promote the extraction of domain-agnostic features. To tightly integrate the domain-agnostic representations and self-supervision, we design a self-generated soft constraint to achieve domain-agnostic contrastive learning (termed "DaCo"). Extensive experiments and analysis on cross-illumination and cross-weather settings are conducted on three challenging datasets. The proposed "DaCo" outperforms current contrastive learning based image retrieval methods by a large margin.