lxc86739795/human_vehicle_parsing_platform

A pytorch codebase for human parsing and vehicle parsing

parsing_platform

A pytorch codebase for human parsing and vehicle parsing.

Introduction

A pytorch codebase for human parsing and vehicle parsing. The introduction of our new MVP dataset for vehicle parsing can be found HERE.

      

Requirements

• Linux or macOS with python ≥ 3.6
• PyTorch = 0.4.1
• torchvision that matches the Pytorch installation. You can install them together at pytorch.org to make sure of this.
• tensorboard (needed for visualization): pip install tensorboard

Supported methods

• PSPNet
• DeepLabV3
• CCNet
• DANet
• OCNet
• CE2P
• HRNet
• BraidNet

Supported datasets

• Look-Into-Person LIP
• Multi-grained Vehicle Parsing MVP

Train and Test

The scripts to train and test models are in train_test. The scripts for PSPNet, DeepLabV3, and HRNet are ready for directly running. The train/val/test splitting files used in our experiments can be found here.

Model Zoo

Models trained on the MVP dataset for vehicle parsing:

Method	Dataset	Pixel Acc	Mean Acc	mIoU	download
PSPNet	MVP-Coarse	90.26%	89.08%	79.78%	model
PSPNet	MVP-Fine	86.21%	69.61%	57.47%	model
DeepLabV3	MVP-Coarse	90.55%	89.45%	80.41%	model
DeepLabV3	MVP-Fine	87.42%	73.50%	61.60%	model
HRNet	MVP-Coarse	90.40%	89.36%	80.04%	model
HRNet	MVP-Fine	86.47%	72.62%	60.21%	model

* The performance is evaluated on the test set.

** The PSPNet and HRNet models are trained with cross-entropy loss. The DeepLabV3 models are trained with cross-entropy + IoU loss.

*** We also released several pre-trained model on the LIP dataset. Please refer to models.

Citation

@inproceedings{mm/LiuZLSM19,
  author    = {Xinchen Liu and
               Meng Zhang and
               Wu Liu and
               Jingkuan Song and
               Tao Mei},
  title     = {BraidNet: Braiding Semantics and Details for Accurate Human Parsing},
  booktitle = ACM MM,
  pages     = {338--346},
  year      = {2019}
}

@inproceedings{mm/LiuLZY020,
  author    = {Xinchen Liu and
               Wu Liu and
               Jinkai Zheng and
               Chenggang Yan and
               Tao Mei},
  title     = {Beyond the Parts: Learning Multi-view Cross-part Correlation for Vehicle
               Re-identification},
  booktitle = {ACM MM},
  pages     = {907--915},
  year      = {2020}
}

human_vehicle_parsing_platform's Issues

Vehicle reid dataste parsing

How to use the trained model HRNet to parse the vehicle reid dateset, such as VeRi776 or Veri-wild

如何解决ImportError: ../../libs/_ext/__ext.so: undefined symbol: state

File "../../libs/_ext/init.py", line 4, in
from .__ext.so import lib as _lib, ffi as _ffi
ImportError: ../../libs/_ext/__ext.so: undefined symbol: state

Help

Hello Doc Liu!I am very interested in this work.When I run the code,I wonder how I can get the pth file of the HRNet pretrained model.Or if you can offer me a MVP trained model?

networks.CE2P has no module Res_Deeplab

human_vehicle_parsing_platform/train_test/DenseASPP/test.py

Line 6 in 524168d

from networks.CE2P import Res_Deeplab

human_vehicle_parsing_platform/train_test/Deeplabv3/test.py

Line 6 in 524168d

from networks.CE2P import Res_Deeplab

Vehicle reid dateset parsing

How to use the trained model HRNet to parse the vehicle reid dateset, such as VeRi776 or Veri-wild. Please help me！

human_vehicle_parsing_platform代码运行后模型问题

您好，我运行了你的human_vehicle_parsing_platform代码报了附件图1的错误，我进行单步调试后发现是附件图2的y(维度8，48，193，193)和self.fuse_layer[0]x维度不一致的问题，self.fuse_layer[0]x维度是（8,48,194,194），附件图3是self.fuse_layer的网络结构，我不知道为什么self.fuse_layer[0]1维度是（8,48,194,194）？这个错误该怎么改呢？在这个之前还遇到一个错误“RuntimeError: cuda runtime error (11) : invalid argument at /opt/conda/conda-bld/pytorch_1535491974311/work/aten/src/THC/THCGeneral.cpp:663”
我看了网上说是2080ti与低版本的pytorch和cuda不匹配问题，所以把cudnn.benchmark改成false，这个错误就消失了，但是遇到了前面提到的维度错误，不知道二者之间是不是有影响？数据集用的你发布的带分割标签的数据集，训练用的coarse_train_id.txt也是你公开的那个。(我运行的是train_test下的Deeplabv3的train.py，里面的run_train.sh运行不起来。config文件和cls_hrnet_w48_sgd_lr5e-2_wd1e-4_bs32_x100.yaml是从BraidNet文件下的复制过来的。)

Question: Inference code

刘博士您好，我想在自己的数据集上使用您的这个项目，请问有可以使用的推理代码吗(. .)