基于多尺度卷积特征融合的行人重识别 
下载： 1647次
Person Reidentification Based on Multiscale Convolutional Feature Fusion
江南大学物联网工程学院物联网应用技术教育部工程中心, 江苏 无锡 214122
图 & 表
图 1. 网络结构
Fig. 1. Network structure
下载图片 查看原文
图 2. 所提分块策略与PCB分块策略比较
Fig. 2. Comparison of our blocking strategy with PCB blocking strategy
下载图片 查看原文
图 3. 二分类的几何示意图。(a)原始的softmax loss;(b)归一化权重和特征后的softmax loss
Fig. 3. Geometric diagram of the second classification. (a) Original softmax loss; (b) softmax loss with normalized weights and features
下载图片 查看原文
图 4. 算法识别效果图。(a) Baseline;(b) PCB (out RPP);(c) baseline+SPP (1+2+6);(d) ours
Fig. 4. Results of recognition algorithm. (a) Baseline; (b) PCB (out RPP); (c) baseline+SPP (1+2+6); (d) ours
下载图片 查看原文
表 1两种方法对基准模型的提升效果
Table1. Effects of two methods for improving baseline model
| Model | Market-1501 | DukeMTMC-reID |
|---|
| rank-1 | mAP | rank-1 | mAP |
|---|
| Baseline | 86.88 | 69.59 | 72.62 | 54.93 | | Baseline+SPP | 89.31 | 71.41 | 80.83 | 64.99 | | Baseline+L2_softmax | 88.72 | 72.42 | 76.93 | 60.18 | | Ours | 91.36 | 75.46 | 82.05 | 65.97 | | Ours+Re-ranking | 93.29 | 88.81 | 86.49 | 81.93 |
|
查看原文
表 2不同分块模型性能比较
Table2. Performance comparison of different block models
| Model | Market-1501 | DukeMTMC-reID |
|---|
| rank-1 | mAP | rank-1 | mAP |
|---|
| Baseline | 86.88 | 69.59 | 72.62 | 54.93 | | PCB(out RPP) | 88.48 | 69.89 | 80.21 | 64.27 | | Baseline+SPP(1+2+3) | 89.31 | 71.41 | 80.83 | 64.99 | | Baseline+SPP(1+2+4) | 89.41 | 70.68 | 79.35 | 62.55 | | Baseline+SPP(1+2+6) | 88.09 | 68.75 | 80.05 | 64.89 |
|
查看原文
表 3本文算法与主流算法的性能比较
Table3. Performance comparison of our algorithm with mainstream algorithms
| Method | Market-1501 | DukeMTMC-reID |
|---|
| rank-1 | mAP | rank-1 | mAP |
|---|
| PSE | 87.7 | 69.0 | 79.8 | 62.0 | | GLAD | 89.9 | 73.9 | — | — | | SVDNet | 82.3 | 62.1 | 76.7 | 56.8 | | MultiScale | 88.9 | 73.1 | 79.2 | 60.6 | | MLFN | 90.0 | 74.3 | 81.0 | 62.8 | | HA-CNN | 91.2 | 75.7 | 80.5 | 63.8 | | DPFL | 88.9 | 72.6 | 79.2 | 60.6 | | Ours | 91.4 | 75.5 | 82.1 | 66.0 | | Ours+Re-ranking | 93.3 | 88.9 | 86.5 | 81.9 |
|
查看原文
徐龙壮, 彭力. 基于多尺度卷积特征融合的行人重识别[J]. 激光与光电子学进展, 2019, 56(14): 141504. Longzhuang Xu, Li Peng. Person Reidentification Based on Multiscale Convolutional Feature Fusion[J]. Laser & Optoelectronics Progress, 2019, 56(14): 141504.
PDF全文
