Language Guided Helmet Object Detection

Helmet object detection is an important task for safety. Existing methods all aim to design a new network to extract better visual features. However, because of the high similarity between helmets and heads, only visual features are not enough. In this paper, we present a novel method, which merges language's semantic information for better performance. Specifically, we first use CLIP to extract the feature vectors of objects. Then, feature vectors are calculated the attention with feature maps from detection models such as YOLOv5 for feature enhancement. Experiments show that our method can get better performance on the SHED and SHDD datasets. Our method improves +1.6% and 2.3% mAP on SHED and SHDD datasets, respectively.