zhuxuhan/one-shot-human-parsing

(AAAI 2021) Progressive One-shot Human Parsing

End-to-end One-shot Human Parsing

This is the official repository for our two papers:

• Progressive One-shot Human Parsing (AAAI 2021)
• End-to-end One-shot Human Parsing (journal version)

---

Introduction:

In the two papers, we propose a new task named One-shot Human Parsing (OSHP). OSHP requires parsing humans in a query image into an open set of reference classes defined by any single reference example (i.e., a support image) during testing, no matter whether they have been annotated during training (base classes) or not (novel classes). This new task mainly aims to accommodate human parsing into a wider range of applications that seek to parse flexible fashion/clothing classes that are not pre-defined in previous large-scale datasets.

Progressive One-shot Human Parsing (AAAI 2021) applies a progressive training scheme and is separated into three stages.

End-to-end One-shot Human Parsing (journal version) is a one-stage end-to-end training method, which has higher performance and FPS.

---

Main results:

You can find the well-trained models together with the performance in the following table.

EOPNet	ATR-OS, Kway F1	ATR-OS, Kway Fold F2	LIP-OS, Kway F1	LIP-OS, Kway F2	CIHP-OS, Kway F1	CIHP-OS Kway F2
Novel mIoU	31.1	34.6	25.7	30.4	20.5	25.1
Human mIoU	61.9	63.3	43.0	45.7	49.1	45.5
Model	Model	Coming Soon	Model	Model	Model	Model

You can find the well-trained models together with the performance in the following table.

EOPNet	ATR-OS, 1way F1	ATR-OS, 1way F2	LIP-OS, 1way F1	LIP-OS, 1way F2	CIHP-OS, 1way F1	CIHP-OS 1way F2
Novel mIoU	53.0	41.4	42.0	46.2	25.4	36.4
Human mIoU	68.2	69.5	57.0	58.0	53.8	55.4
Model	Coming Soon

---

Getting started:

Data preparation:

First, please download ATR, LIP and CIHP dataset from source. Then, use the following commands to link the data into our project folder. Please also remember to download the atr flipped labels and cihp flipped labels.

# ATR dataset
$ ln -s YOUR_ATR_PATH/JPEGImages/* YOUR_PROJECT_ROOT/ATR_OS/trainval_images
$ ln -s YOUR_ATR_PATH/SegmentationClassAug/* YOUR_PROJECT_ROOT/ATR_OS/trainval_classes
$ ln -s YOUR_ATR_PATH/SegmentationClassAug_rev/* YOUR_PROJECT_ROOT/ATR_OS/Category_rev_ids


# LIP dataset
$ ln -s YOUR_LIP_PATH/TrainVal_images/TrainVal_images/train_images/* YOUR_PROJECT_ROOT/LIP_OS/trainval_images
$ ln -s YOUR_LIP_PATH/TrainVal_images/TrainVal_images/val_images/* YOUR_PROJECT_ROOT/LIP_OS/trainval_images
$ ln -s YOUR_LIP_PATH/TrainVal_parsing_annotations/TrainVal_parsing_annotations/train_segmentations/* YOUR_PROJECT_ROOT/LIP_OS/trainval_classes
$ ln -s YOUR_LIP_PATH/TrainVal_parsing_annotations/TrainVal_parsing_annotations/val_segmentations/* YOUR_PROJECT_ROOT/LIP_OS/trainval_classes
$ ln -s YOUR_LIP_PATH/Train_parsing_reversed_labels/TrainVal_parsing_annotations/* YOUR_PROJECT_ROOT/LIP_OS/Category_rev_ids
$ ln -s YOUR_LIP_PATH/val_segmentations_reversed/* YOUR_PROJECT_ROOT/LIP_OS/Category_rev_ids


# CIHP dataset
$ ln -s YOUR_CIHP_PATH/Training/Images/* YOUR_PROJECT_ROOT/CIHP_OS/trainval_images
$ ln -s YOUR_CIHP_PATH/Validation/Images/* YOUR_PROJECT_ROOT/CIHP_OS/trainval_images
$ ln -s YOUR_CIHP_PATH/Training/Category_ids/* YOUR_PROJECT_ROOT/CIHP_OS/trainval_classes
$ ln -s YOUR_CIHP_PATH/Validation/Category_ids/* YOUR_PROJECT_ROOT/CIHP_OS/trainval_classes
$ ln -s YOUR_CIHP_PATH/Category_rev_ids/* YOUR_PROJECT_ROOT/CIHP_OS/Category_rev_ids

Please also download our generated support .pkl files from source, which contains each class's image IDs. You can also generate support files on your own by controlling dtrain_dtest_split in oshp_loader.py, however, the training and validation list might be different from our paper.

Finally, your data folder should look like this:

${PROJECT ROOT}
|-- data
|   |--datasets
|       |-- ATR_OS
|       |   |-- list
|       |   |   |-- meta_train_id.txt
|       |   |   `-- meta_test_id.txt
|       |   |-- support
|       |   |   |-- meta_train_atr_supports.pkl
|       |   |   `-- meta_test_atr_supports.pkl
|       |   |-- trainval_images
|       |   |   |-- 997-1.jpg
|       |   |   |-- 997-2.jpg
|       |   |   `-- ...
|       |   |-- trainval_classes
|       |   |   |-- 997-1.png
|       |   |   |-- 997-2.png
|       |   |   `-- ... 
|       |   `-- Category_rev_ids
|       |       |-- 997-1.png
|       |       |-- 997-2.png
|       |       `-- ... 
|       |-- LIP_OS
|       |   |-- list
|       |   |   |-- meta_train_id.txt
|       |   |   |-- meta_test_id.txt
|       |   |-- support
|       |   |   |-- meta_train_lip_supports.pkl
|       |   |   `-- meta_test_lip_supports.pkl
|       |   |-- trainval_images
|       |   |   |-- ...
|       |   |-- trainval_classes
|       |   |   |-- ... 
|       |   `-- Category_rev_ids
|       |       |-- ... 
|       `-- CIHP_OS
|           |-- list
|           |   |-- meta_train_id.txt
|           |   |-- meta_test_id.txt
|           |-- support
|           |   |-- meta_train_cihp_supports.pkl
|           |   `-- meta_test_cihp_supports.pkl
|           |-- trainval_images
|           |   |-- ...
|           |-- trainval_classes
|           |   |-- ... 
|           `-- Category_rev_ids
|               |-- ... 

Finally, please download the DeepLab V3+ pretrained model (pretrained on COCO dataset) from source and put it into the data folder:

${PROJECT ROOT}
|-- data
|   |--pretrained_model
|       |--deeplab_v3plus_v3.pth

Installation:

Please make sure your current environment has Python >= 3.7.0 and pytorch >= 1.1.0. The pytorch can be downloaded from source.

Then, clone the repository and install the dependencies from the following commands:

git clone https://github.com/Charleshhy/One-shot-Human-Parsing.git
cd One-shot-Human-Parsing
pip install -r requirements.txt

Training:

To train EOPNet in End-to-end One-shot Human Parsing (journal version), run:

# OSHP kway on ATR-OS fold 1
bash scripts/atr_eop_kwf1.sh

Validation:

To evaluate EOPNet in End-to-end One-shot Human Parsing (journal version), run:

# OSHP kway on ATR-OS fold 1
bash scripts/evaluate_atr_kwf1.sh

---

TODO:

• Release training/validation code for POPNet
• Release well-trained EOPNet 1-way models

Citation:

If you find our papers or this repository useful, please consider cite our papers:

@inproceedings{he2021progressive,
title={Progressive One-shot Human Parsing},
author={He, Haoyu and Zhang, Jing and Thuraisingham, Bhavani and Tao, Dacheng},
booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
year={2021}
}

@article{he2021end,
title={End-to-end One-shot Human Parsing},
author={He, Haoyu and Zhang, Jing and Zhuang, Bohan and Cai, Jianfei and Tao, Dacheng},
journal={arXiv preprint arXiv:2105.01241},
year={2021}
}

Acknowledgement:

This repository is mainly developed basing on Graphonomy and Grapy-ML.