基于任务关联的端到端目标检测算法

Task Aligned End-to-End Object Detection

对预测结果准确排序是端到端目标检测的关键。已有的端到端检测器将分类和定位当作独立任务,减少了两者之间的关联,导致利用分类分数排序的结果不可靠,降低了检测性能。针对上述问题,本文从样本选择、损失函数和网络结构三个方面进行了优化,提高了两者之间的一致性。首先利用分类和定位的排序结果计算样本选择的代价矩阵,并优先考虑分类和定位一致性大的样本作为正样本。另外,使用基于任务关联的损失函数训练分类器,学习同时表示目标分类精度和定位准确度的分数。考虑到分类和定位对特征需求的差异,在头部检测网络中引入特征对齐层,缓解了分类和定位之间的冲突。在COCO数据集上,基于任务关联的端到端目标检测算法的性能优于许多优秀的检测器。

Accurate ranking of prediction results is the key to end-to-end object detection. Existing end-to-end detectors treat classification and localization as independent tasks, reducing the correlation between them, and resulting the unreliable results ranking by classification scores, which reduces the detection performance. To solve the above problem, we try to improve the correlation between classification and localization form label assignment, loss function and network structure. First, the cost matrix of sample selection is calculated by using the sorting results of classification and location, and the samples with high consistency of classification and location are given priority as positive samples. We propose a task-aligned loss function to train the classifier, which aims to learn a score that can simultaneously represent the accuracy of object classification and localization. We introduce a feature alignment layer in the head detection network, which alleviates the conflict between classification and localization at the feature extraction level. On COCO datasets, the end-to-end object detection algorithm proposed in this paper outperforms many excellent detectors.