HRM-CenterNet: A High-Resolution Real-time Fittings Detection Method

Most successful fittings detectors are anchor-based, which is challenging to meet the lightweight and real-time requirements of the edge computing system. We propose a high-resolution real-time network HRM-CenterNet. Firstly, the lightweight MobileNetV3 is used to extract multi-level features from images. Then, to improve the resolution of the feature maps and reduce the spatial semantic information loss during the image downsampling process, a high-resolution feature fusion network based on iterative aggregation is introduced. Finally, we conduct experiments on the PASCAL VOC dataset and fittings dataset. The results show that HRM-CenterNet improves accuracy as well as robustness, and meets the performance requirements of real-time edge detection.