基于二次生成对抗的人体姿态估计 下载: 890次
Human Pose Estimation Based on Secondary Generation Adversary
贵州大学大数据与信息工程学院大数据与智能技术重点实验室, 贵州 贵阳 550025
图 & 表
图 1. 本方法的结构示意图
Fig. 1. Structure schematic diagram of our method
下载图片 查看原文
图 2. heatmap示意图
Fig. 2. Schematic diagram of heatmap
下载图片 查看原文
图 3. 沙漏网络结构
Fig. 3. Structure of hourglass network
下载图片 查看原文
图 4. SHN的级联结构图
Fig. 4. Cascade structure diagram of SHN
下载图片 查看原文
图 5. 中间监督结构
Fig. 5. Structure of intermediate supervision
下载图片 查看原文
图 6. ASR流程
Fig. 6. Procedure of ASR
下载图片 查看原文
图 7. AHO流程
Fig. 7. Procedure of AHO
下载图片 查看原文
图 8. heatmap重建
Fig. 8. Reconstruction of heatmap
下载图片 查看原文
图 9. 不同方法得到的heatmaps。(a)文献[
4];(b)文献[
10]; (c)文献[
11];(d)本方法
Fig. 9. Heatmaps obtained by different methods. (a) Ref. [4]; (b) Ref. [10]; (c) Ref. [11]; (d) ours
下载图片 查看原文
图 10. 关节点估计误差对比
Fig. 10. Comparison of joint estimation errors
下载图片 查看原文
表 1批次图像训练流程
Table1. Training process of batch images
Input: a mini-batch training image set X |
---|
1.X is randomly and equally divided into X1、X2、X3;2.Train D1 using X1;3.Train G1、D1 using X2 with table 2 on ASR;4.Train G1、D1 using X3 with table 2 on AHO. |
|
查看原文
表 2单图像训练流程
Table2. Training process of single image
Input: image x |
---|
1.Get shortcut features from D1;2.Get distribution P from shortcut features in G1;3.Sample an adversarial augmentation data from P;4.Compute the loss of D1: with ;5.Random augment x to get ;6.Compute the loss of D1: with ;7.Compare and with formula (5) and formula (6) to update G1;8.Update D1. |
|
查看原文
表 3第二次生成对抗的训练流程
Table3. Training process of the secondary generation adversary
Input: image x;ground truth heatmap C |
---|
1. D2 reconstructs heatmap: D(C,x);2. Compute Lreal with formula (11);3. G2 generates predictive heatmap: =G(x);4. Compute LMSE with formula (8);5. D2 reconstructs heatmap:D(,x);6. Compute ;7. Compute Lfake、L 'D with formula (11)、formula (12);8. Update D2;9. Compute Ladv、LG with formula (9)、formula (10);10.Update G2. |
|
查看原文
表 4不同方法在LSP数据集的PCK
Table4. PCK of different methods in LSP data setunit: %
Method | Head | Shoulder | Elbow | Wrist | Hip | Knee | Ankle | Mean |
---|
Ref. [21] | 97.8 | 92.5 | 87.0 | 83.9 | 91.5 | 90.8 | 89.9 | 90.5 | Ref. [4] | 98.2 | 94.0 | 91.2 | 87.2 | 93.5 | 94.5 | 92.6 | 93.0 | Ref. [12] | 98.5 | 94.0 | 89.8 | 87.5 | 93.9 | 94.1 | 93.0 | 93.1 | Ref. [10] | 98.6 | 95.3 | 92.8 | 90.0 | 94.8 | 95.3 | 94.5 | 94.5 | Ref. [11] | 98.2 | 94.9 | 92.2 | 89.5 | 94.2 | 95.0 | 94.1 | 94.0 | Ours | 98.8 | 95.7 | 92.6 | 90.8 | 94.8 | 96.1 | 95.0 | 94.8 |
|
查看原文
表 5不同方法在MPII数据集的PCKh
Table5. PCKh of different methods in the MPII data setunit: %
Method | Head | Shoulder | Elbow | Wrist | Hip | Knee | Ankle | Mean |
---|
Ref. [21] | 97.8 | 95.0 | 88.7 | 84.0 | 88.4 | 82.8 | 79.4 | 88.5 | Ref. [4] | 98.2 | 96.3 | 91.2 | 87.1 | 90.1 | 87.4 | 83.6 | 90.9 | Ref. [12] | 98.6 | 96.2 | 90.9 | 86.7 | 89.8 | 87.0 | 83.2 | 90.6 | Ref. [10] | 98.1 | 96.6 | 92.5 | 88.4 | 90.7 | 87.7 | 83.5 | 91.5 | Ref. [11] | 98.2 | 96.8 | 92.2 | 88.0 | 91.3 | 89.1 | 84.9 | 91.8 | Ours | 98.4 | 97.1 | 93.4 | 88.7 | 92.5 | 90.3 | 85.2 | 92.2 |
|
查看原文
表 6模型使用效率的对比
Table6. Comparison of model efficiency
Method | Convergenceiteration times | Average processingtime /s | GFLOPs /(109 times) | Number ofparameters /107 |
---|
Ref. [11] | 19500 | 0.48 | 10.820 | 5.495 | Ours | 26600 | 0.73 | 13.702 | 6.738 |
|
查看原文
张显坤, 张荣芬, 刘宇红. 基于二次生成对抗的人体姿态估计[J]. 激光与光电子学进展, 2020, 57(20): 201509. Xiankun Zhang, Rongfen Zhang, Yuhong Liu. Human Pose Estimation Based on Secondary Generation Adversary[J]. Laser & Optoelectronics Progress, 2020, 57(20): 201509.