Abstract:

Person re-identification is a hot topic in computer community and has many applications, such as transportation, public safety and video surveillance. This paper proposes a new method, which incorporates Camera Style Learning（CSL） and Multi-Granularity Loss（MGL）, has achieved state-of-the-art results in person re-ID domain. CSL can reduce the impact caused by the difference of camera style, which can preferably play the advantage of triplet loss and effectively improve the recognition accuracy. MGL is adopted in the learning process, by which multiple levels of feature maps are exploited to make the learned features more discriminate. Experiments on two large-scale datasets, Market-1501 and DukeMTMC-reID, show the proposed method obtains 3.7% and 3.2% improvement in Rank-1 precision over the baseline Camstyle and the results reach 93.2% and 81.5%, respectively. On the Market-1501 dataset, this paper achieves 93.8% in mAP and 96.1% in Rank-1 accuracy after re-ranking, after MGL is applied, surpassing the state of the art.

Key words: camera style learning, triplet loss, person re-identification, multi-granularity loss

摘要：

行人重识别是计算机领域的一个热门话题，在交通、公共安全和视频监控等场景有着广泛的应用。提出了摄像头风格学习（CSL）结合多粒度损失（MGL）的新方法，在行人重识别领域取得了优势性能。通过摄像头风格学习可以减少由摄像头差异带来的影响，更好地发挥triplet loss的优势，有效地提高识别精度。在学习过程中结合多粒度损失，利用多个层次的特征图，使学习到的特征更有区分力。在Market-1501和DukemMTMC-reID两个大型数据集上做了对比实验，实验结果表明，提出的方法优于原Camstyle方法，在Rank1上提高了3.7%和3.2%，准确率分别达到93.2%和81.5%。在Market-1501数据集上结合多粒度损失并使用re-ranking方法后，Rank1的准确率为96.1%，mAP的准确率为93.8%，获得了当前已发表最高准确度。

关键词: 摄像头风格学习, triplet loss, 行人重识别, 多粒度损失