.png){kind=icon}
.png){kind=icon}
Return
PCGDNet: Reliability-aware glass detection with physics-inspired transparent-surface consistency from monocular RGB images
.png){kind=icon}
PRE
AI
T
J
J
H
W
H
DOI：10.1016/j.eswa.2026.133893.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Glass detection from monocular RGB images remains challenging because glass regions often exhibit weak boundaries, limited intrinsic texture, and appearances shaped by transmitted background content, reflections, highlights, and blur. These factors can result in incomplete segmentation and false responses in complex scenes. To address this problem, we propose PCGDNet, a reliability-aware and consistency-guided framework for monocular RGB glass detection. The framework first estimates ambiguity-prone regions through a failure-aware RGB encoding module and produces reliability-aware feature embeddings. A Refraction-Transmission Consistency-Guided Module then captures observable image-domain cues of transparent surfaces, including interior continuity and boundary-local disturbance, without estimating explicit physical quantities such as refractive index, surface normal, or transmission coefficient. An uncertainty-guided feature modulation module allocates appearance, boundary, contextual, and consistency-related cues according to local reliability, while a refinement module improves region completeness and boundary quality. On the GDD benchmark, PCGDNet achieves an MAE of 0.032, an IoU of 92.78, an Fβ-measure of 0.961 on glass images, and an FPR of 0.12 on non-glass images. Evaluations on GSD and an RGB-only transparent-object subset further suggest transferability to related transparent-surface scenarios under RGB-only input. The method may still be affected by extreme reflection, severe illumination changes, very low-contrast boundaries, and unusual transparent-surface geometries. These results indicate the effectiveness of reliability-aware cue allocation and image-domain transparent-surface consistency for monocular RGB glass detection.
Keywords:
Glass detection
Monocular RGB images
Transparent-surface perception
Reliability-aware perception
Image-domain transparent-surface consistency
Uncertainty-guided fusion
Journal
IF:
7.5
Papers:
2.9W
Citations:
10.2W
Organization
Cited Papers