Meta-YOLO: Meta-Learning for Few-Shot Traffic Sign Detection via Decoupling Dependencies

Abstract

Considering the low coverage of roadside cooperative devices at the current time, automated driving should detect all road markings relevant to driving safety, such as traffic signs that tend to be of great variety but are fewer in number. In this work, we propose an innovative few-shot object detection framework, namely Meta-YOLO, whose challenge is to generalize to the unseen classes by using only a few seen classes. Simply integrating the YOLO mechanism into a meta-learning pipeline will encounter problems in terms of computational efficiency and mistake detection. Therefore, we construct a two-stage meta-learner ℱ model that can learn the learner initialization, the learner update direction and learning rate, all in a single meta-learning process. To facilitate deep networks with learning, the fidelity features of the targets improve the performance of meta-learner ℱ, but we also design a feature decorrelation module (FDM), which firstly transforms non-linear features into computable linear features based on RFF, and secondly perceives and removes global correlations by iteratively saving and reloading the features and sample weights of the model. We introduce a three-head module to learn global, local and patch correlations with the category detection result outputted by the aggregation in meta-learner ℱ, which endows a multi-scale ability with detector ϕ. In our experiments, the proposed algorithm outperforms the three benchmark algorithms and improves the mAP of few-shot detection by 39.8%.