Joint Discriminative and Generative Learning for Person Re-identification

Person re-identification (re-id) remains challenging due to significant intra-class variations across different cameras. Recently, there has been a growing interest in using generative models to augment training data and enhance the invariance to input changes. The generative pipelines in existing methods, however, stay relatively separate from the discriminative re-id learning stages. Accordingly, re-id models are often trained in a straightforward manner on the generated data. In this paper, we seek to improve learned re-id embeddings by better leveraging the generated data. To this end, we propose a joint learning framework that couples re-id learning and data generation end-to-end. Our model involves a generative module that separately encodes each person into an appearance code and a structure code, and a discriminative module that shares the appearance encoder with the generative module. By switching the appearance or structure codes, the generative module is able to generate high-quality cross-id composed images, which are online fed back to the appearance encoder and used to improve the discriminative module. The proposed joint learning framework renders significant improvement over the baseline without using generated data, leading to the state-of-the-art performance on several benchmark datasets.

PDF Abstract CVPR 2019 PDF CVPR 2019 Abstract

Results from the Paper


Task Dataset Model Metric Name Metric Value Global Rank Result Benchmark
Person Re-Identification CUHK03 DG-Net MAP 61.1 # 12
Rank-1 65.6 # 11
Person Re-Identification DukeMTMC-reID DG-Net(RK) Rank-1 90.26 # 29
mAP 88.31 # 17
Person Re-Identification DukeMTMC-reID DG-Net Rank-1 86.6 # 50
mAP 74.8 # 53
Unsupervised Person Re-Identification DukeMTMC-reID->Market-1501 DGNet mAP 26.83 # 6
Rank-1 56.12 # 5
Rank-10 72.18 # 3
Rank-5 78.12 # 2
Unsupervised Person Re-Identification DukeMTMC-reID->MSMT17 DGNet mAP 6.35 # 5
Rank-1 20.59 # 4
Rank-10 31.67 # 3
Rank-5 37.04 # 2
Unsupervised Domain Adaptation Duke to Market DG-Net mAP 26.83 # 21
rank-1 56.12 # 21
rank-5 72.18 # 15
rank-10 78.12 # 15
Unsupervised Domain Adaptation Duke to MSMT DG-Net mAP 6.35 # 10
rank-1 20.59 # 10
rank-5 31.67 # 9
rank-10 37.04 # 10
Person Re-Identification Market-1501 DG-Net Rank-1 94.8 # 60
mAP 86.0 # 69
Person Re-Identification Market-1501 DG-Net(RK) Rank-1 95.4 # 45
mAP 92.49 # 22
Person Re-Identification Market-1501-C DG-Net Rank-1 31.75 # 12
mAP 9.96 # 15
mINP 0.35 # 8
Unsupervised Person Re-Identification Market-1501->DukeMTMC-reID DGNet mAP 24.25 # 5
Rank-1 42.62 # 5
Rank-10 64.63 # 3
Rank-5 58.57 # 3
Unsupervised Person Re-Identification Market-1501->MSMT17 DG-Net mAP 5.41 # 5
Rank-1 17.11 # 5
Rank-10 26.66 # 3
Rank-5 31.62 # 2
Unsupervised Domain Adaptation Market to Duke DG-Net mAP 24.25 # 20
rank-1 42.62 # 20
rank-5 58.57 # 14
rank-10 64.63 # 14
Unsupervised Domain Adaptation Market to MSMT DG-Net mAP 5.41 # 11
rank-1 17.11 # 11
rank-5 26.66 # 10
rank-10 31.62 # 11
Person Re-Identification MSMT17 DG-Net Rank-1 77.2 # 27
mAP 52.3 # 26
Rank-5 87.4 # 1
Rank-10 90.5 # 1
Unsupervised Person Re-Identification MSMT17->DukeMTMC-reID DGNet Rank-1 61.89 # 2
Rank-10 75.81 # 2
Rank-5 80.34 # 2
mAP 40.69 # 2
Unsupervised Person Re-Identification MSMT17->Market-1501 DG-Net Rank-1 61.76 # 2
Rank-10 83.25 # 2
Rank-5 77.67 # 2
mAP 33.62 # 2
Person Re-Identification UAV-Human DG-Net mAP 61.97 # 2
Rank-1 65.81 # 1
Rank-5 85.71 # 1

Methods