Automatic label assignment object detection mehtod on only one feature map
Tingsong Ma,Zengxi Huang,Nijing Yang,Changyu Zhu,Ping Deng
DOI: https://doi.org/10.1007/s00138-023-01481-4
IF: 2.983
2023-11-07
Machine Vision and Applications
Abstract:Most deep learning-based object detection methods are proposed based on multi-level feature environments. Although some researchers have tried to detect on one-level features, where multiple feature maps are utilized. In this paper, we aim to propose a novel anchor-free object detection approach with an automatic label assignment strategy on only one feature map. The proposed method follows the main idea of AutoAssign to achieve the label assignment strategy. However, to make the strategy work appropriately in one feature map environment, several modifications have been made. A prior knowledge fusion method called ‘Center Weighting Fusion’ is proposed for label assignment strategy, where the Gaussian mixture model function is applied to calculate the weights of each object on one feature map. By doing so, some objects that are close to each other, whose weights will be merged and generate some points (Recheck Points) that are shared with multiple objects. For those ‘Recheck Points,’ the detector will judge how many objects share this point and generate a corresponding number of different-sized proposals. In detecting different-sized objects, we propose a ‘Uniform Detection’ method to limit each point’s regression distance according to the target’s category. A large number of experimental data show that the proposed method presents competitive detection accuracy with normal anchor-free detectors (43.8% mAP), while it is smaller (30% smaller) and faster (50% better).
computer science, cybernetics, artificial intelligence,engineering, electrical & electronic