基于HPLF的行人再识别

为了更好的挖掘局部特征,提升行人再识别的精度,本文提出了一种利用水平池化提取局部特征的HPLF(Horizontal Pooling for Local Feature)算法,在ResNet-50网络中对输入的联合数据集进行预处理,提取特征,对ResNet-50网络生成的特征图进行水平切割,通过分割的特征图计算两两特征之间的距离,再用难样本三元组损失(Triplet loss with Hard example mining, TriHard loss)来作为局部特征损失函数训练,通过特征图计算全局距离,通过难样本三元组损失来训练,将这两个损失函数加上一个Softmax交叉熵损失函数,联合起来作为总的损失函数进行参数修正.实验结果表明:在Market1501数据集中, mAP (mean Average Precision), Rank-1, Rank-5, Rank-10等性能指标上, HPLF算法比其他算法有3%左右的提升.

Pedestrian Re-Identification Based on HPLF

Pedestrian re-identification generally considered as a sub-problem of image retrieval. Due to the distance between the camera and the pedestrian, the definition of the pedestrian photo is generally fuzzy, and the camera’s view angle of pedestrians is fixed, so it is not enough to recognize pedestrians by faces. In order to better mine strong local features and improve the accuracy of pedestrian re-identification, this study proposes an algorithm, namely Horizontal Pooling for Local Feature(HPLF). We preprocess the input joint data set in ResNet-50 network, extract features, and horizontally cut the feature map generated by ResNet-50 network, with which we calculate the distance between every two features. Triple loss with hard example mining(TriHard loss) is used for training as a local feature loss function.The global distance is calculated according to the feature map and trained through TriHard loss. The two loss functions plus a Softmax cross entropy loss function are combined as the total loss function for parameter correction. The experimental results show that HPLF’s performances of mean Average Precision(mAP), Rank-1, Rank-5, and Rank-10 in the Market1501 data set are about 3% higher than those of other algorithms.