All Projects → abeardear → FashionAI-Keypoint

abeardear / FashionAI-Keypoint

Licence: other
fashionAI clothes keypoint detection

Programming Languages

python
139335 projects - #7 most used programming language
shell
77523 projects

Projects that are alternatives of or similar to FashionAI-Keypoint

eval-mpii-pose
Evaluation code for the MPII human pose dataset
Stars: ✭ 58 (+205.26%)
Mutual labels:  pose-estimation
openpose-docker
A docker build file for CMU openpose with Python API support
Stars: ✭ 68 (+257.89%)
Mutual labels:  pose-estimation
LoFTR
Code for "LoFTR: Detector-Free Local Feature Matching with Transformers", CVPR 2021
Stars: ✭ 1,046 (+5405.26%)
Mutual labels:  pose-estimation
TransPose
PyTorch Implementation for "TransPose: Keypoint localization via Transformer", ICCV 2021.
Stars: ✭ 250 (+1215.79%)
Mutual labels:  pose-estimation
DeepLabCut-core
Headless DeepLabCut (no GUI support)
Stars: ✭ 29 (+52.63%)
Mutual labels:  pose-estimation
label-studio-frontend
Data labeling react app that is backend agnostic and can be embedded into your applications — distributed as an NPM package
Stars: ✭ 230 (+1110.53%)
Mutual labels:  pose-estimation
slamkit
SLAM Kit
Stars: ✭ 28 (+47.37%)
Mutual labels:  pose-estimation
LEMO
Official Pytorch implementation for 2021 ICCV paper "Learning Motion Priors for 4D Human Body Capture in 3D Scenes" and trained models / data
Stars: ✭ 149 (+684.21%)
Mutual labels:  pose-estimation
ICON
ICON: Implicit Clothed humans Obtained from Normals (CVPR 2022)
Stars: ✭ 641 (+3273.68%)
Mutual labels:  pose-estimation
awesome-6d-object
Awesome work on object 6 DoF pose estimation
Stars: ✭ 252 (+1226.32%)
Mutual labels:  pose-estimation
HybrIK
Official code of "HybrIK: A Hybrid Analytical-Neural Inverse Kinematics Solution for 3D Human Pose and Shape Estimation", CVPR 2021
Stars: ✭ 395 (+1978.95%)
Mutual labels:  pose-estimation
sleap
A deep learning framework for multi-animal pose tracking.
Stars: ✭ 200 (+952.63%)
Mutual labels:  pose-estimation
MSPN
Multi-Stage Pose Network
Stars: ✭ 321 (+1589.47%)
Mutual labels:  pose-estimation
cvxpnpl
A Perspective-n-Points-and-Lines method.
Stars: ✭ 56 (+194.74%)
Mutual labels:  pose-estimation
Fast Stacked Hourglass Network OpenVino
A fast stacked hourglass network for human pose estimation on OpenVino
Stars: ✭ 52 (+173.68%)
Mutual labels:  pose-estimation
mediapipe-osc
MediaPipe examples which stream their detections over OSC.
Stars: ✭ 26 (+36.84%)
Mutual labels:  pose-estimation
articulated-pose
[CVPR 2020, Oral] Category-Level Articulated Object Pose Estimation
Stars: ✭ 85 (+347.37%)
Mutual labels:  pose-estimation
sc depth pl
Pytorch Lightning Implementation of SC-Depth (V1, V2...) for Unsupervised Monocular Depth Estimation.
Stars: ✭ 86 (+352.63%)
Mutual labels:  pose-estimation
HRFormer
This is an official implementation of our NeurIPS 2021 paper "HRFormer: High-Resolution Transformer for Dense Prediction".
Stars: ✭ 357 (+1778.95%)
Mutual labels:  pose-estimation
Multi-Person-Pose-using-Body-Parts
No description or website provided.
Stars: ✭ 41 (+115.79%)
Mutual labels:  pose-estimation

fashionAI 服装关键点检测

天池算法大赛服装关键点检测,给定五种类型的服装,采用人体姿态估计的方法检测关键点。最终结果排名24,CPN模型没有复现很好,略遗憾...

0.效果预览

1.模型

模型结合了HourGlass模型和CPN模型,其中HG堆叠了2个,另外在HG上采样过程的1/2大小的特征图上添加了热点图监督标签。

RGB图像送入两个分支网络分别计算,最后concat二者的特征图,具体结构如图所示。

添加了soft-argmax层,可以由热点图转化到具体的坐标值,用坐标值groundtruth监督学习

2.策略

最多只能使用两个不同参数的模型,检测模型也算。通过检测可以提高目标占比,提升效果。

使用第一级预测结果截取目标,为了防止截取不完整,向外扩展30像素,再训练第二级crop模型。

第一级模型testB线上4.17%,crop之后的模型testB线上4.05%,融合之后3.95%.

3.训练与预测细节

  • 优化器Adam,学习率1e-3,每过10个epoch乘0.1,CPN的残差网络部分,学习率为全局学习率的0.1倍。

  • 初始化,HG用了开源的一个人体姿态估计参数初始化,CPN残差部分用Imagenet预训练参数。

  • 高斯热点图的亮块大小为13x13,输入图像512,输出热点图大小128

  • 数据增强,颜色变换,旋转变换

  • 图片扩展至512而不是直接resize,左右居中。

  • 预测时,正向和水平翻转各预测一遍,再平均结果

4.数据准备

训练数据集fashionAI_key_points_train_20180227.tar.gz和[update] warm_up_train_20180222.tar解压后的Image一起放在data/train中,并将两个训练数据集的train.csv合并成trainnew.csv,再从中划分出验证集valid.csv

测试数据fashionAI_key_points_test_b_20180418.tgz解压后放入testb文件夹,应该放入的是Images文件夹和test.csv

HG模型初始化使用了MPII人体关键点数据预训练

5.运行环境

pytorch 0.3.0

visdom 用于训练损失曲线可视化

6.程序运行

程序运行为.sh脚本执行,將会训练模型,对testb做预测并生成提交结果.

train.sh 执行训练程序,每次训练5个模型,并保存在checkpoint文件夹

submit.sh 执行测试提交程序,將生成提交文件.csv

python demo.py 可以看两级预测结果与最后的融合结果

chmod +x ./train.sh
./train.sh

chmod +x ./submit.sh
./submit.sh
Note that the project description data, including the texts, logos, images, and/or trademarks, for each open source project belongs to its rightful owner. If you wish to add or remove any projects, please contact us at [email protected].