Transparency detection in autonomous driving

Transparent objects are difficult for visual perception systems because their appearance depends strongly on the surrounding scene. A glass surface can transmit background texture, reflect nearby objects, and still have weak object boundaries. This makes glass detection relevant for autonomous-driving perception, where transparent obstacles may otherwise be interpreted as free space. This thesis investigates whether synthetic data generated in CARLA can be used to adapt an existing glass detection model to autonomous-driving scenes. A data generation pipeline was built to render RGB images, segmentation masks, and metadata for glass planes placed in road scenes. Two datasets were generated. The first dataset uses transparent glass planes and empty negative samples. The second dataset uses zero-thickness glass planes and opaque hard negatives to reduce the risk that the model learns simple border, thickness, or plane-shape cues. The pretrained GDNet model from Don’t Hit Me! Glass Detection in Real-World Scenes was first evaluated directly on the generated CARLA test sets. It performed poorly in this zero-shot setting, reaching an IoU of 0.116 on the standard glass-plane dataset and 0.063 on the thin-plane dataset. The low precision values show that the model produced many false-positive glass predictions in the synthetic driving domain. After fine-tuning on the generated CARLA training data, performance improved strongly. The respective fine-tuned models reached an IoU of 0.979 on the standard dataset and 0.983 on the thin-plane dataset. On the thin-plane test set, the model also produced no positive predictions on opaque hard-negative or empty negative samples. This suggests that the model did not simply classify every plane-like object as glass. The results show that synthetic CARLA data can be used to adapt a pretrained glass detection model to a controlled autonomous-driving domain. At the same time, the evaluation remains limited to synthetic test data. Real-world validation is required before drawing conclusions about transfer to real driving scenes.