Traffic Sign Recognition and Retrieval Using Limited Dataset in the Wild
Enke Li, Lingrui Mei · 2021
Most deep learning models require a large amount of annotated data for training, many individuals and institutions have provided a large number of high-quality open datasets, and quite a lot of academic works has been done based on such open datasets and solved quite a lot of classical problems in the field of computer vision. However, the actual scenarios of many engineering problems and the scenarios covered by open datasets are very different, and the data distribution of the data taken from different scenarios varies greatly, so it is difficult for models trained from open datasets to solve the problems encountered in actual scenarios. In addition, the cost of model training and data labeling is relatively high. For an already trained classification or model, adding new data often requires retraining, and the newly added data also needs to be labeled according to certain labeling standards, which greatly increases the monetary cost and time cost of using deep learning models in industry. In this paper, we propose a object detection process that can identify unlabeled categories of traffic signals: we use a object detection model to extract traffic labels based on an image content retrieval method that partially trains the labeled data to match the traffic sign categories that do not appear in the training data. The method is tested on the TSRD dataset and achieves functions that are not possible with traditional end-to-end detectors, while the proposed method also achieves good results on object detection tasks that are achievable with object detectors.