2021青海省首届河湟杯数据湖算法大赛—行人精细化识别赛道

写在前面：今年2月左右关注和更进的一个比赛，尝试整合过一些行人属性识别的文章trick，由于近期临近毕业事情比较多，该比赛延期后不打算继续更进了，分享一些尝试过的代码和想法。

Introduction

该比赛是接近标准的行人属性识别任务，但是有些互斥类和二分类两种形式的标注，二分类（穿着颜色属性）中是软标签。Baseline尝试过两个Repo:
- https://github.com/chufengt/iccv19_attribute (想法不错但精度上尚有提升空间)
- https://github.com/valencebond/Rethinking_of_PAR （Trick和参数设置得很好，精度较高）
本项目在Rethinking_of_PAR基础上进行了实验和一些补充：
- 数据增强
  - Random rotate/Erase
  - Multi-Scale Traning/Test
- 尝试的Backbone：
  - ResNet
  - CoatNet
  - ConvNext
  - Vit
  - swin-transformer
- 尝试的Neck
  - ALM (个人的实现方式) [1]
  - FPN (常规FPN)
- 尝试的Head
  - Linear Layer (wi/o BN)
  - CSRA [2]
  - MSSC [3]

[1] Improving Pedestrian Attribute Recognition With Weakly-Supervised Multi-Scale Attribute-Specific Localization (ICCV2019)

[2] Residual Attention: A Simple but Effective Method for Multi-Label Recognition (ICCV2021)

[3] Improving Pedestrian Attribute Recognition with Multi-Scale Spatial Calibration (IJCNN)

欢迎 star和交流

Name		Name	Last commit message	Last commit date
Latest commit History 2 Commits
configs		configs
dataset		dataset
losses		losses
metrics		metrics
models		models
optim		optim
scheduler		scheduler
tools		tools
PARInfer.py		PARInfer.py
PARInfer2.py		PARInfer2.py
PARTrain.py		PARTrain.py
PARval.py		PARval.py
Process.txt		Process.txt
README.md		README.md
VARInfer.py		VARInfer.py
batch_engine.py		batch_engine.py
infer.py		infer.py
requirements.txt		requirements.txt
result.csv		result.csv
result_par_res50_ema.csv		result_par_res50_ema.csv
result_par_swin_b_csra.csv		result_par_swin_b_csra.csv
result_par_swin_fpnpooling_csra.csv		result_par_swin_fpnpooling_csra.csv
result_par_swin_l_csra8.03.csv		result_par_swin_l_csra8.03.csv
result_swin_b.csv		result_swin_b.csv
train.py		train.py
train_gpu.sh		train_gpu.sh
train_gpu_par.sh		train_gpu_par.sh
train_gpu_var.sh		train_gpu_var.sh