VDT-RealWorld

The dataset consists of 324 RGB images accompanied by their corresponding thermal and depth maps. To guarantee broad diversity and comprehensive coverage, we curate the dataset by integrating real-world tri-modal panoramic segmentation datasets with V\u2013T datasets that contain a variety of adverse weather conditions, where the D modality for the latter is synthesized by a monocular depth estimation network. Salient objects were identified through a consensus process among multiple independent reviewers, ensuring their prominence in RGB, thermal, and depth modalities. Pixel-level ground truth masks for SOD were subsequently annotated by professional annotators.