基于多配置特征包的目标检测

针对目标检测中的特征失配问题，提出多配置特征包的概念，刻画同一特征可能出现的不同失配情况。目标分类器学习时，利用Boosting算法学习出最具鉴别力特征包，每个特征包对应一个单特征和它的失配情况，目标分类器是最优特征包分类器的线性组合。进一步地，引入多示例学习思想，有效评估特征包鉴别力、学习特征包分类器。在人脸数据集上的实验表明，较之传统方法，考虑特征失配后，文中算法能获得更好的检测性能。同时，与固定包生成方式相比，多配置特征包能较好拟合特征失配情况，在提高检测率的同时获得更小的检测器尺寸。

Object Detection Based on Multi-Configuration Feature Bag

To solve the problem of feature misalignment appearing in object detection, a multi-configuration feature bag is put forward and used to describe variant misalignments for the same feature. By boosting algorithm, the most discriminative feature bags, which consist of single features and their corresponding misalignment cases, are selected in the phase of classifier training. Based on those best feature bags, weak classifiers are generated and combined into the final ensemble classifier. Moreover, multiple instance learning is introduced to efficiently evaluate the discriminative ability of feature bags and train feature bag classifiers. The experimental results on public face dataset demonstrate that the proposed algorithm is more robust than the traditional method when the problem of feature misalignment is considered. Furthermore, compared to the feature bag with fixed size, the proposed multi-configuration feature bag models the feature misalignment better, gets smaller detector size while improving the detection accuracy.