Traffic sign detection using YOLO: Evaluating failures on localized data

Traffic sign detection (TSD) is an ongoing significant challenge to the implementation of fully self-driving (FSD) technology. While TSD has been extensively studied on standardized benchmark databases such as GTSDB and TT100K, efforts to localize models on regional traffic sign data such as in Indonesia has been lacking. This study evaluates currently available TSD dataset in Indonesia by training an image detection model; You Only Look Once (YOLO) on publicly available localized data and then evaluating the model on new traffic sign images collected from Google Street View in Indonesia. The model reached an impressive [email protected] of 97.07%, however it performed poorly on real-world images, only reaching 28.68% accuracy. Qualitative analysis reveals major issues in current datasets: low variation of images, heavy augmentation inflating true image count, class imbalance improper labelling, and only representing 40 classes out of the 350 officially defined in Indonesia. Future research on this topic should focus on gathering a more varied dataset and integrating specialized modules such as OCR to handle numeric and text signs.