Breaking the Dataset Shackles: Data-E cient Learning with Mamba Network for 360° Salient Object Detection

Abstract 360° salient object detection (SOD) is crucial for analyzing and understanding panoramic scenes. However, current 360° SOD methods face a significant limitation: models are often optimized for individual datasets, leading to dataset-specific overfitting and poor generalization in downstream applications. To overcome this challenge, we propose a data-efficient cross-dataset collaborative training strategy. By integrating four 360° SOD datasets and addressing redundancy and label conflicts, we construct a refined training corpus that retains 66.7% of the original data volume. To fully leverage multi-source data while minimizing spherical distortion artifacts, we introduce 360Mamba—the first Mamba-based network designed specifically for 360° scenes. Experiments show that our method outperforms state-of-the-art models on two benchmark datasets with 33% less training data, demonstrating the benefits of multi-dataset collaborative training and the Mamba architecture.