Image change detection is a fundamental problem for robotic map maintenance and long-term map learning. Local feature-based image comparison is one of the most basic schemes for addressing this problem. However, the local-feature approach encounters difficulties when the query and reference images involve different domains (e.g., time of the day, weather, season). In this paper, we address the local-feature approach from the novel perspective of object-level region features. This study is inspired by the recent success of object-level region features in cross-domain visual place recognition (CD-VPR). Unlike the previous contributions of the CD-VPR task, in the cross-domain change detection (CD-CD) tasks, we consider matching a small part (i.e., the change) of the scene and not the entire image, which is considerably more demanding. To address this issue, we explore the use of two independent object proposal techniques: supervised object proposal (e.g., YOLO) and unsupervised object proposal (e.g., BING). We combine these techniques and compute appearance features of their arbitrarily shaped objects by aggregating local features from a deep convolutional neural network (DCN). Experiments using a publicly available cross-season NCLT dataset validate the efficacy of the proposed approach.
Graph-based local feature adaptation for cross-domain person re-identification
