【摘要】 视频行人重识别在监控场景中起着非常重要的作用.但是，大多数现有方法没有充分利用行人视频序列的时空信息.具体来说，这些方法以相同的分辨率和网络结构处理每一帧图像，造成连续帧特征的高度相似.此外，现有方法通常通过引入各种复杂的操作提高精度，过多的计算开销使其不利于真实场景的部署.针对上述问题，本文提出了一个时序多尺度互补网络，旨在高效地为视频的连续帧提取互补的特征.具体来说，时序多尺度互补网络包含多个具有不同输入分辨率的分支.其中，高分辨率分支处理原始分辨率帧，用于保留行人的细节线索；低分辨率分支处理以不同降采样率得到的低分辨率帧，用于捕捉更全局的行人信息.通过将连续帧输入到不同分支中，连续帧能关注不同粒度的空间区域，生成互补的特征.进一步，设计了一个多分支批量归一化层，保证了训练时分支之间的互补性.最后，提出一个跨分支融合模块，将低分辨率分支的全局信息逐步传播到高分辨分支中，得到一个融合了多尺度全局粗粒度和局部细粒度互补信息的特征.在iLIDS-VID,MARS和LS-VID三个数据集上的实验显示，本文提出的方法达到了比目前最好方法更好的性能，例如，在LS-VID上提升了4.5%mAP和3.1%top-1精度，证明了该方法的有效性.此外，通过降低输入帧的分辨率和使用更小的网络处理低分辨率帧，本文方法大幅度降低了计算开销，仅需要大多数现有方法约35%的计算开销.更多还原

【Abstract】 Person re-identification(reID) aims to match the same person across multiple non-overlapped cameras, which plays an important role in surveillance video analysis. Recently, with the emergence of large video benchmarks and the growth of computational resource, video-based person reID has been attracting a lot of attention. However, most existing methods do not make full advantage of rich spatial and temporal clues in the videos. To be specific, the consecutive frames of a pedestrian video are highly similar,but the existing methods conduct the same operation with the same input resolution on each frame. As a result, the existing methods typically produce highly redundant features for consecutive frames. The redundant frames easily focus on the same most representative local part, which may be difficult to distinguish the persons with seemingly similar local part. In addition, it is common for recent methods to improve the accuracy by introducing more operations, which is not conducive to deployment in many real scenes. In this paper, we present a powerful and efficient video person reID network, Temporal Multi-Scale Complementary Network(TMSCN), to extract complementary features for consecutive frames of a video efficiently. Specifically, TMSCN is built on a multi-branch architecture where each branch has a specific input resolution. High-resolution Branch processes frames at original resolution to preserve the detailed visual clues, and Low-resolution Branches with different down-sampling rates are employed to progressively enlarge the receptive field for capturing global information. By feeding consecutive frames into different branches, TMSCN can enforce consecutive frames to focus on the regions with different spatial scales. Such that the diverse visual features can be discovered for consecutive frames and finally form an integral characteristic of the target identity. Furthermore, since low-resolution images contain fewer details, a small network is sufficient to extract its discriminative features. While high-resolution images contain richer details and require a larger network to process. So TMSCN uses network structures of different capacities for different branches, which can reasonably allocate the computational resources.TMSCN also shares the parameters of the same stage for different branches, which needs no extra parameters over single-branch video re ID networks. Furthermore, we design a Multi-Branch Batch Normalization(MBBN) which uses separated mean and variance for BN layer on each branch. MBBN can ensure stable training and maintain the complementarity among branches. At last, a Cross-Branch Fusion Module(CBPM) is proposed to successively propagate the global and coarse features of low-resolution branch to high-resolution branch. In this way, the final multi-scale feature contains both local details and global complementary information, showing stronger discrimination ability. Experiments on three benchmarks, MARS, LS-VID and i LIDS-VID, show that the proposed method achieves state-of-the-art performance. Specifically, it improves the performance of existing video reID methods by about 4.5% mean Average Precision(mAP) and 3.1% top-1 accuracy on LS-VID benchmark, validating the effectiveness of the proposed method. Moreover, by down-sampling some frames to low-resolution and using small network for low-resolution frames, TMSCN greatly reduces the computations, requiring about 65% less computation cost than most existing methods.更多还原