.png){kind=icon}
Return
Input-robust depth completion for transparent objects
.png){kind=icon}
PRE
AI
DOI：10.1007/s00530-026-02683-2.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Transparent objects are significantly challenging for RGB-D depth sensors due to their non-Lambertian optical properties, resulting in incomplete or erroneous depth measurements. Existing deep learning methods for transparent object depth completion exhibit excessive dependency on raw depth map inputs, making them vulnerable to two critical failure modes: (1) sensor domain shifts when deployed with different depth cameras, and (2) performance degradation under depth map perturbations. To address these limitations, we propose RIDCNet (Robust Input-Dependent Depth Completion Network), which reduces reliance on raw depth by explicitly incorporating RGB-derived geometric features through surface normal estimation. Our architecture decouples feature extraction into two parallel branches: RGB-to-Normal for extracting scale-invariant geometric cues from RGB images, and Depth-to-Depth for processing raw depth. A cross-task fusion module is introduced to integrate complementary information between the two branches. Extensive experiments on benchmark datasets demonstrate that RIDCNet achieves state-of-the-art performance while exhibiting superior robustness: maintaining competitive accuracy under synthetic depth perturbations and achieving improved cross-sensor generalization compared to previous methods.
Keywords:
Depth estimation
Transparent objects
Depth sensors
Journal
IF:
3.1
Papers:
2.8K
Citations:
2.7K
Organization
Cited Papers
ShareSaveShareSaveShareSavePAR-mono: monocular video depth estimation network based on channel separation and dynamic attention.png)
MULTIMEDIA SYSTEMS
MULTIMEDIA SYSTEMSIF3.1
PREAI
ShareSaveSRSA-Depth: shape and region similarity awareness for outdoor monocular depth estimationMULTIMEDIA SYSTEMS
MULTIMEDIA SYSTEMSIF3.1
PREAI
ShareSave
View more