基于视角信息嵌入的行人重识别 
下载： 1263次
Person Re-Identification Based on View Information Embedding
哈尔滨工程大学信息与通信工程学院, 黑龙江 哈尔滨 150001
图 & 表
图 1. PSE网络模型结构
Fig. 1. Structure of PSE network model
下载图片 查看原文
图 2. 基于视角信息嵌入模型结构
Fig. 2. Model based perspective information embedding
下载图片 查看原文
图 3. 深度可分离卷积结构
Fig. 3. Depthwise separable convolution
下载图片 查看原文
图 4. 改进的深度可分离卷积结构
Fig. 4. Improved depthwise separable convolution
下载图片 查看原文
图 5. 深度可分离模块结构
Fig. 5. Structure of depthwise separable module
下载图片 查看原文
表 1视角预测模块有效性验证实验结果
Table1. Results of perspective predictor module verification experiment
| Method | Market1501 | Duke-MTMC-reID | MARS | | |
|---|
| rank-1 /% | mAP /% | rank-1 /% | mAP /% | rank-1 /% | mAP /% |
|---|
| Proposed | Except perspective | 83.6 | 62.6 | 74.1 | 53.7 | 67.7 | 50.1 | | All | 89.9 | 71.6 | 79.9 | 61.7 | 74.1 | 57.6 |
|
查看原文
表 2改进深度可分离卷积的有效性验证实验结果
Table2. Results of improved depthwise separable convolution verification experiment
| Method | Market1501 | Duke-MTMC-reID | MARS | | |
|---|
| rank-1 /% | mAP /% | rank-1 /% | mAP /% | rank-1 /% | mAP /% |
|---|
| Proposed | Except SE-Block | 87.0 | 67.5 | 77.8 | 59.6 | 71.2 | 54.2 | | All | 89.9 | 71.6 | 79.9 | 61.7 | 74.1 | 57.6 |
|
查看原文
表 3中层特征方法的有效性验证实验结果
Table3. Verification experiment results of mid-level feature method
| Method | Market1501 | Duke-MTMC-reID | MARS | | |
|---|
| rank-1 /% | mAP /% | rank-1 /% | mAP /% | rank-1 /% | mAP /% |
|---|
| Proposed | Except Mid-level-feature | 87.4 | 70.9 | 79.5 | 57.8 | 72.1 | 55.9 | | All | 89.9 | 71.6 | 79.9 | 61.7 | 74.1 | 57.6 |
|
查看原文
表 4改进模型有效性验证实验结果
Table4. Results of improved model verification experiment
| S/N | Improved method | Market1501 | Duke-MTMC-reID | MARS | | | |
|---|
| IM1 | IM2 | IM3 | rank-1 /% | mAP /% | rank-1 /% | mAP /% | rank-1 /% | mAP /% |
|---|
| PSE | ① | - | - | - | 87.7 | 69.0 | 79.8 | 62.0 | 72.1 | 56.9 | | ② | - | - | Y | 86.8 | 66.5 | 77.5 | 60.5 | 70.5 | 54.9 | | ③ | - | Y | - | 82.5 | 64.2 | 70.1 | 54.3 | 68.4 | 49.3 | | ④ | - | Y | Y | 85.3 | 65.8 | 72.5 | 57.2 | 70.2 | 53.8 | | Ours | ⑤ | Y | - | - | 87.5 | 69.0 | 79.4 | 61.1 | 70.9 | 57.1 | | ⑥ | Y | - | Y | 85.9 | 66.5 | 75.6 | 59.9 | 67.8 | 54.6 | | ⑦ | Y | Y | - | 86.6 | 65.6 | 77.4 | 59.4 | 70.1 | 55.1 | | ⑧ | Y | Y | Y | 89.9 | 71.6 | 79.9 | 61.7 | 74.1 | 57.6 |
|
查看原文
表 5算法运行速度对比实验结果
Table5. Results of algorithm running speed comparison experiment
| Method | Time /s |
|---|
| Total match | Per match (19720) |
|---|
| PSE | 141.57 | 0.0072 | | Proposed | 288.96 | 0.0147 |
|
查看原文
表 6算法结果对比
Table6. Comparison of algorithm results
| Method | Market1501 | Duke-MTMC-reID | MARS | | |
|---|
| rank-1 /% | mAP /% | rank-1 /% | mAP /% | rank-1 /% | mAP /% |
|---|
| P2S(point to set) | 70.7 | 44.3 | - | - | - | - | | Spindle | 76.9 | - | - | - | - | - | | Consistent aware | 80.9 | 55.6 | - | - | - | - | | GAN(generative adversarial networks) | 78.1 | 56.2 | 67.7 | 47.1 | - | - | | Latent parts | 80.3 | 57.5 | - | - | 71.8 | 56.1 | | ResNet+OIM(online instance matching) | 82.1 | - | 68.1 | - | - | - | | ACRN(attribute-complementary re-ID net) | 83.6 | 62.6 | 72.6 | 52.0 | - | - | | SVD(singular value decomposition) | 82.3 | 62.1 | 76.7 | 56.8 | - | - | | Part aligned | 81.0 | 63.4 | - | - | - | - | | PDC(pose-driven deep convolutional model) | 84.1 | 63.4 | - | - | - | - | | JLML(jointly learning multi-loss) | 85.1 | 65.5 | - | - | - | - | | DPFL | 88.6 | 72.6 | 79.2 | 60.6 | - | - | | Forest | - | - | - | - | 70.6 | 50.7 | | DGM(dynamic graph matching)+IDE | - | - | - | - | 65.2 | 46.8 | | QMA | - | - | - | - | 73.7 | 51.7 | | ResNet baseline | 82.6 | 59.8 | 71.5 | 50.3 | 64.5 | 49.5 | | PSE | 87.7 | 69.0 | 79.8 | 62.0 | 72.1 | 56.9 | | Proposed algorithm | 89.9 | 71.6 | 79.9 | 61.7 | 74.1 | 57.6 |
|
查看原文
毕晓君, 汪灏. 基于视角信息嵌入的行人重识别[J]. 光学学报, 2019, 39(6): 0615007. Xiaojun Bi, Hao Wang. Person Re-Identification Based on View Information Embedding[J]. Acta Optica Sinica, 2019, 39(6): 0615007.
PDF全文
