[论文笔记]CVPR2017_Person Re-identification in the Wild

Title: Person Re-identification in the Wild

Authors: Liang Zheng¹, Hengheng Zhang², Shaoyan Sun³, Manmohan Chandraker⁴, Yi Yang¹, Qi Tian²

Affiliations: 

1. University of Technology Sydney
2. UTSA
3. USTC
4. UCSD & NEC Labs

两组大佬同期发表关于person search的文章（另一篇），虽然这篇题目是person re-id。person reid发展到了这个时期，detection+re-id作为两个分开的任务已经需要更进一步发展了，即将两个任务融合在一起，更符合现实性。XIAO Tong等提出了一个dataset，并提出了端到端的person search方法；这篇文章，Zheng Liang大大也提出了PRW数据集，同时提出了一种端到端的 person reid的pipeline。

---

Contribution

• A novel large-scale dataset, Person Re-identification in the Wild (PRW), for simultaneous analysis of person detection and re-ID.
• Comprehensive benchmarking of state-of-the-art detection and recognition methods on the PRW dataset.
• Novel insights into how detection aids re-ID, along with an effective fine-tuning strategy and similarity measure to illustrate how they might be utilized.
• Novel insights into the evaluation of pedestrian detectors for the specific application of person re-ID.

The PRW Dataset

• 6个摄像头：5 are 1080×1920 HD and 1 is 576×720 SD.

• 训练集 5, 704 frames and 482 IDs

• 测试集 6, 112 frames and 450 IDs

• 评价指标：

  • we mainly use the precision-recall curve and average precision to evaluate detection performance

  • also report the log-average miss rate (MR)

• 2,057 query images for the 450 IDs in the test set, averaging 4.57 (maximum 6) queries/ID

• the gallery in PRW varies with the settings of pedestrian detectors

• 用IoU评价detection，用mAP和rank-1,10,20 accuracy评价re-id

Pipeline of an end-to-end person re-ID system

Proposed Improvements

Cascaded fine-tuning strategy

包含两种，前一种是前人的，后一种是作者提出的

Confidence Weighted Similarity

之前的文章在比较query和gallery images时都将所有gallery boxes视作同等重要，这样会导致一个问题：当gallery的规模增大时，不可避免地会出现误检的结果，导致精度下降。因此作者将detection confidence和similarity mearsurement相乘，从直觉上表示false positive的目标的权重相应减小。

Experiments