A state-space model-based approach for transparent object depth estimation

Abstract Transparent object recognition and perception remain challenging tasks. To address the limitations of depth sensors in accurately capturing the surface depth of transparent objects, this paper introduces a state-space model-based network for transparent object depth estimation. The network processes RGB images and raw depth data, leveraging the state-space model’s global feature extraction capability to capture multi-scale features and improve object-background differentiation. To enhance RGB-D multi-modal fusion, an advanced feature fusion module is designed to establish feature correlations by integrating visual cues from RGB images with geometric cues from depth maps. Experiments on the ClearGrasp dataset demonstrate that the proposed network significantly outperforms existing methods and exhibits strong robustness in real-world scenarios.