This work presents a traffic sign recognition system built on YOLOv2 that adds early-exit branches and geometric features. The branches allow the network to stop early for easy cases, reducing computation time. The geometric features use an unsupervised Bayesian segmentation to create binary sign templates, which are compared with detected bounding boxes to correct classification errors. Experiments on a dataset combining GTSDB and GTSRB with 3,000 training and 300 test images across ten sign classes show that the branched architecture alone matches baseline YOLOv2 accuracy (0.680 mAP) with slightly faster runtime (0.647 s vs 0.6607 s). Adding geometric verification during inference raises mAP to 0.713, and using geometric features during training achieves 0.697 mAP. The work is interesting because it demonstrates that a simple geometric post-processing step can improve classification accuracy without significantly slowing inference, and that early exits can speed up detection for easy cases.
Branched YOLOv2 with geometric features improves traffic sign recognition speed and accuracy.
The authors show that early termination for easy cases and geometric template matching reduce inference time and classification errors.
Academic
Arefeh Rezaei
K. N. Toosi University of Technology
Research Digest··2 min read
In one line
Branched YOLOv2 with geometric features improves traffic sign recognition mAP from 0.680 to 0.713.
Read it as
The authors extended YOLOv2 with intermediate prediction layers (branched architecture) to allow early exit for easy cases, and introduced geometric features from Bayesian segmentation to distinguish visually similar signs.
The authors constructed a dataset by combining GTSDB and GTSRB using seamless cloning and controlled image transformations.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.