GlassFormer: Real-time Glass Segmentation using Radar-Depth Fusion

Suhani Grover¹    Astik Srivastava¹    Viswas Dinesh¹    Avinash Sharma²    K. Madhava Krishna¹   

¹ Robotics Research Center, IIIT Hyderabad, India    ² IIT Jodhpur, India   

Transparent surfaces are a persistent failure case for robotic perception: RGB cameras see the background behind glass, and depth sensors return invalid or background measurements at transparent interfaces. GlassFormer fuses 60.5 GHz millimeter-wave radar with RGB-D sensing. Radar reflects strongly off glass precisely where vision and depth fail, giving a geometric cue that is insensitive to lighting. We turn the radar–depth inconsistency into a coarse spatial prior and inject it into a SegFormer-B2 backbone via cross-modal RadarAttention, achieving 0.88 mIoU on a mixed-condition split and 0.59 mIoU on a dedicated low-light split, in real time on resource-constrained platforms.