Self-Correction-Human-Parsing-on-Binary-Images

Original Images	Contrast Enhancement

Output Pascal Model	Binary Image-Input Model	Inference Result

• train.py
• dataset.ipynb
• dataset_and_train.ipynb
• simple_extractor.py(inference)
• requirements.txt
• evaluate.py

Dataset

Dataset name: bodies-at-rest

Github repository: https://github.com/Healthcare-Robotics/bodies-at-rest

We used real dataset of bodies-at-rest dataset that contains 20 human participants (10M/10F) with 1K labeled real pressure images.

• prepairing binary images dataset:

For preparing dataset, we used pascal pretrained model that implemented for RGB images. The Pascal Person Part has 7 labels, including 'Background', 'Head', 'Torso', 'Upper Arms', 'Lower Arms', 'Upper Legs', 'Lower Legs'.

The steps for preparing the database are as follows

1- Extraction of image field of dataset pickle file

2- Improve image contrast

3- Create segmented images using pascal model

4- Convert segmented images using pascal model to binary images

5- Creating labels of binary images that contains 7 classes

For prepairing binary images dataset, you can clone this repositiry and run dataset.ipynb file.

step1	step2	step3	step4	step5

• In this step you need trained models to prepaire dataset. If you have trouble downloading models, you can download manually using this link: Trained Models

Instalation

1- Clone this repository using the following command:

https://github.com/NahidEbrahimian/Human-Parsing-on-Binary-Image

2- In ./Human-Parsing-on-Binary-Images directory, run the following command to install requirements:

pip install -r requirements.txt

Train

1- Clone this repository using the following command:

https://github.com/NahidEbrahimian/Human-Parsing-on-Binary-Image

There are two solutions for training:

1- You can run dataset_and_train.ipynb file for both prepairing dataset and training

2- Afther prepairing dataset, run the following command:

%cd ./Human-Parsing-on-Binary-Image
!python train.py --data-dir ./dataset/dataset --num-classes 7 --batch-size 3 --imagenet-pretrain ./pretrain_model/resnet101-imagenet.pth

• Used wandb in training pocess.

Inference

1- For inference, first download pretrained model on binary images from this link and put in ./log directory: download model

2- Put your binary images in ./input directory

3- Run the following command:

!python simple_extractor.py --dataset 'pascal' --model-restore "./log/checkpoint_40.pth.tar" --input-dir './input' --output-dir './output'