TGADHead: An efficient and accurate task-guided attention-decoupled head for single-stage object detection

In object detection, localization and classification of the targets are two fundamental subtasks that underpin the application of many knowledge-based intelligent models in various visual fields. However, current methods show inconsistency between the two subtasks, i.e., accurate localization may show a poor classification score or vice versa, due to inadequate design of the detectors. This inconsistency significantly degrades the overall detection accuracy. To address this issue, we propose a novel Task-Guided Attention-Decoupled Head (TGADHead) to improve detection accuracy using an efficient single-stage approach. The proposed framework consists of two inter-connected components: Task Decoupled Attention Distributor (TDAD) and Task Correlation Network (TCN). In the first component, we propose TDAD with two well-designed task specific attention perceptrons to enhance the spatial information required for localization and the semantic information required for classification, respectively. This task specific prediction mechanism improves classification and location performance. Secondly, we construct a Task Correlation Network (TCN) for each of decoupled branch to transfer the correlation between classification and localization features while maintaining the consistency of the two subtasks, i.e., an accurate object prediction outcome should simultaneously have a high quality location boundary box and a high classification score. We achieve +1.9 Average Precision (AP) on the MS-COCO compared to the state-of-the-art detectors.