Skip to content

对论文Attention-A-Lightweight-2D-Hand-Pose-Estimation-Approach-master进行测试和Pytorch复现。

Notifications You must be signed in to change notification settings

hanchenchen/Attention-A-Lightweight-2D-Hand-Pose-Estimation-Approach-master

Repository files navigation

Attention-A-Lightweight-2D-Hand-Pose-Estimation-Approach-master TEST

任务描述

在手部关键点检测任务中,对论文 Attention! A Lightweight 2D Hand Pose Estimation Approach 中提出的Attention Augmented Inverted Bottleneck Block等结构进行测试。 Pytorch版本:https://github.com/hanchenchen/Attention-A-Lightweight-2D-Hand-Pose-Estimation-Approach-Pytorch/tree/main

测评环境

  • Ubuntu 16.04.6 LTS
  • Python 3.8.5
  • tensorflow 2.4.1

数据集

[CMU Panoptic][ http://domedb.perception.cs.cmu.edu/handdb.html] [SHP][https://sites.google.com/site/zhjw1988/] [FreiHAND Dataset][https://lmb.informatik.uni-freiburg.de/resources/datasets/FreihandDataset.en.html] [HO3D_v2][https://cloud.tugraz.at/index.php/s/9HQF57FHEQxkdcz?]
31836 36000 130240 66034

所有图片剪裁为224*224大小,部分数据集的3D关键点标注投影为2D。

训练样本:验证样本:测试样本 = 80%:10%:10%

测评指标

PCKProbability of Correct Keypoint within a Normalized Distance Threshold

测试大纲

使用Convolutional Pose Machines(CPM)作为参考的基准,测试论文中提出的architecture是否有良好的性能

采用消融实验方法,对论文中使用的 Attention Augmented Inverted Bottleneck Block、Blur (Pooling Method)、Mish(Activation Function)进行测试。

在原项目的基础上添加如下代码:

train.py: 增加了parser和json配置文件,便于在多个数据库上进行训练。

evaluate.py: 使用PCK指标对模型进行量的测试和质的测试,结果存放在文件夹qualitative_results、quantitative_results。

(dataset_path)/crop_images.py: 将不同数据集中的图片剪裁为特定大小(224),并对labels进行修改

(dataset_path)/make_tfrecord.py: 将不同的数据集制作为tfrecord文件

model_ablation.py + arch.json: 实现了 IV. EVALUATION - B. Ablation studies 中的12种 architectures

model_cpm:使用Convolutional Pose Machines作为基准。

pck.py: 计算PCK。

print_logs.py: 打印训练日志(loss,acc,pck)

compare.py: 比较不同模型的PCK结果。

Train
python train.py (datatset_name) --arch (1-12/cpm) --GPU 0
python train.py HO3D_v2 --arch 1 --GPU 0
Evaluate
python evaluate.py (datatset_name) --arch (1-12/cpm) --GPU 0

测试结果

Ablation Study

在HO3D_v2数据集上,对CPM,Arch1、2、3、4 一共5个模型进行训练,取20个Epoch中val_loss最小的模型进行比较。

  • CPM:baseline, Total params: 15,987,291
  • Arch1:Attention module:1,Pooling Method:Blur, Total params: 1,970,674
  • Arch2:Attention module:0,Pooling Method:Blur, Total params: 1,072,850
  • Arch3:Attention module:0,Pooling Method:Average, Total params: 1,072,850
  • Arch4:Attention module:1,Pooling Method:Average, Total params: 1,970,674

Datasets

Architecture1 在不同数据集上的表现,Epoch = 15, 取val_loss最优模型。

结果分析

  • 论文提出的结构相较于CPM更加Lightweight。
  • Arch1 的准确率仍然和CPM有较大的差距,考虑如下原因:
    • CPM使用了Heatmap,有利于坐标的学习。论文提出的结构没有使用Heatmap。
  • Arch1 与 Arch2 进行比较,添加了 Self-Attention 结构后反而PCK下降,考虑了如下原因:
    • 原论文中使用了SGD优化器,而 SGD 的缺点在于收敛速度慢,可能在鞍点处震荡。这可能导致了Arch1的loss达到0.06之后便难以下降。
  • Blur Pooling 使有 Self-Attention 结构的Arch1 表现优于Arch4;但在无 Self-Attention 结构的Arch2、3中,与Average Pooling 表现相似。
Weights

https://www.dropbox.com/sh/99u7apw2q52mzn2/AAD0JAmOQ8P4ZK-8VDXDR6xqa?dl=0

Reference

[1] Santavas N, Kansizoglou I, Bampis L, et al. Attention! a lightweight 2d hand pose estimation approach[J]. IEEE Sensors Journal, 2020. [[code]][https://github.com/nsantavas/Attention-A-Lightweight-2D-Hand-Pose-Estimation-Approach]

[2] Chen Y, Ma H, Kong D, et al. Nonparametric structure regularization machine for 2D hand pose estimation[C]//Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. 2020: 381-390. [code][https://github.com/HowieMa/NSRMhand]

[3] Wei S E, Ramakrishna V, Kanade T, et al. Convolutional pose machines[C]//Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. 2016: 4724-4732.

[4] Simon T, Joo H, Matthews I, et al. Hand keypoint detection in single images using multiview bootstrapping[C]//Proceedings of the IEEE conference on Computer Vision and Pattern Recognition. 2017: 1145-1153. [Panoptic][http://domedb.perception.cs.cmu.edu/handdb.html]

[5] Zimmermann C, Ceylan D, Yang J, et al. Freihand: A dataset for markerless capture of hand pose and shape from single rgb images[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision. 2019: 813-822. [FreiHAND][https://lmb.informatik.uni-freiburg.de/projects/freihand/]

[6] Zhang J, Jiao J, Chen M, et al. 3d hand pose tracking and estimation using stereo matching[J]. arXiv preprint arXiv:1610.07214, 2016. [SHP]

[7] Shivakumar S H, Oberweger M, Rad M, et al. HO-3D: A Multi-User, Multi-Object Dataset for Joint 3D Hand-Object Pose Estimation[J]. arXiv. org e-Print archive, 2019. [HO3D_v2]

About

对论文Attention-A-Lightweight-2D-Hand-Pose-Estimation-Approach-master进行测试和Pytorch复现。

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages