Task-aware Disentanglement for Object Detection

Sibling-head structure is widely used to alleviate the feature conflict between classification and regression tasks in most object detectors. However, as the two branches of the sibling head are trained with exactly the same positive samples and lack explicit feature disentanglement in the forward propagation, the classification-sensitive features and localization-sensitive features are still somewhat coupled. As a result, the feature conflict between the two tasks still remains, which seriously hurts the performance of the classifier and regressor in the testing phase. In this paper, we propose a Task-Aware Disentangled object Detector (TDD) that explicitly disentangles the classification and regression from the aspect of feature disentanglement and sampling strategy. In terms of feature disentanglement, we design a task-aware activation head driven by a reconstruction-activation mechanism to explicitly activate corresponding sensitive features for classification and localization in the forward propagation. Furthermore, we explore a novel task-aware sampling strategy that explicitly assigns the task-adaptive samples for classification and regression tasks according to their quality distributions. Extensive experiments on MS COCO show that our TDD consistently surpasses the baseline by similar to 2.0 AP with different backbones. Moreover, our best model achieves 55.1 AP, outperforming most state-of-the-art detectors.