Iterative search & learn for sign detection in large datasets
Gary Overett, Wei Wang · 2017
Leading methods for object detection and recognition using neural networks improve markedly when given very large training sets. In the case of infrequent objects, such as traffic signs, hand labeling many hours of road scene video becomes impractical. In this paper, we propose an iterative Search & Learn method capable of quickly creating object detection datasets numbering in the 10's of thousands of real-world examples within days, using only a few hours of human labelling. The approach can be initiated using a few hundred hand labeled objects or a single exemplar. This is used to create a “search-detector” with sufficient precision to find a larger collection of real-world examples from a large video library. These examples can then be used to train an improved search-detector. A similar approach can be simultaneously applied to traffic sign classification, where a preliminary classifier can be used to minimize human sorting in the creation of very large classification datasets. Results are shown on 3 traffic sign detection and classification tasks; speed-limit (10 subclasses), minimum-speed (3 subclasses), and no-entry. The resulting datasets contain over 200K, 100K, and 80K real-world examples respectively in 3-5 iterations of the method. Classification accuracy rapidly improves with the number of real-world examples and generally approaches 99% accuracy once the number of objects surpasses 100K images.