TransGeo6D: A Lightweight Geometry-Aware Framework for Transparent Object Depth Completion and 6-DoF Pose Candidate Generation

Transparent objects produce incomplete RGB-D measurements, degrading scene geometry and downstream pose-candidate generation. We designed TransGeo6D, a lightweight two-stage framework linking transparent-object depth completion to six-degree-of-freedom pose-candidate generation. Lite-DCNet combines explicit depth-validity and gradient cues with learned RGB-D features through geometry-aware gating and convolutional gated recurrent refinement. Lite-GSNet integrates a Ghost-based sparse backbone with channel- and point-wise candidate attention. Candidate-point pooling yields permutation-invariant channel descriptors, while twofold rotational symmetry identifies equivalent orientations of the ideal parallel-jaw representation. These properties support feature aggregation and the interpretation of candidate-pose consistency. On the scene-exclusive TransCG test set, Lite-DCNet has 0.308 million parameters and achieves a mean absolute depth error of 10.26 mm. Lite-GSNet has 2.3953 million parameters and reaches 15.19 frames per second in module-level profiling. With Lite-GSNet frozen on 30 scene-disjoint TransCG test scenes, Lite-DCNet achieves translation and rotation drifts of 57.95 mm and 39.77°, respectively, with a PCR@30/30 of 27.40% and a GT-CF@50 of 33.62%, the highest rates among the three compared learned depth front ends. Controlled regional perturbations reveal spatially nonuniform sensitivity to comparable depth disturbances, linking depth recovery to candidate stability under the parallel-jaw orientation equivalence.