Trans3-Vision: Transfer Learning Based Transformer for Transparent Object Segmentation with Grounded-ID2

Autonomous driving presents a captivating future application of AI. Experimenting with SOTA commercially available autonomous vehicles and with the autonomous vehicle we developed, we discovered that recognizing transparent objects like glass remains a significant challenge for autonomous systems. Modern architecture frequently uses glass, which LiDAR systems in autonomous vehicles struggle to detect due to insufficient reflection, leading to potential collisions. Transparent objects also mislead systems that use cameras. This issue limits the application of autonomous driving technology in environments with extensive glass usage. This issue hinders the application of autonomous driving technology around these new environments. Therefore, in this research, we developed a training method for a machine vision AI model to detect transparent objects. During the research process, we primarily addressed the following challenges: •We developed an automatic dataset generation methodology and pipeline and Grounded-SAM Integrated Diffusion In-painting Dataset (Grounded-ID2), a real-world-simulating glass segmentation dataset. •Proposed an improved Transfer learning-based Transformer for the Transparent Object Segmentation model (Trans3-Vision) to overcome the performance losses trained with combined datasets. The process introduced by Grounded-ID2 is not limited to glass and transparent object segmentation and applies to all other areas where image masks are needed.