We introduce a dual-head structure designed to jointly handle change classification and cross-visibility confidence estimation.
Chamelion is a novel approach for online change detection and long-term 3D map management, leveraging a dual-head network and composition-based augmentation for robust generalization across sensors and environments.
Training change detection requires multi-session data, which is costly and impractical to collect. To address this, we propose a single-session augmentation method that synthetically generates change pseudo-labels using only a single traversal of the environment.
We introduce a dual-head structure designed to jointly handle change classification and cross-visibility confidence estimation.
The class head predicts whether each point is a change or static, while the confidence head estimates the probability that the point is visible in both the map and scan.
We evaluate the performance of each method in both scan-wise and map-wise change detection.
We qualitatively evaluated the generalization performance across various 3D range sensors, as shown in below.
The model was trained solely on our pseudo-labeled dataset, without additional training on other datasets, demonstrating the proposed data generation method's effectiveness and applicability across different environments and 3D range sensors.