secretstar112 / mmsegmentation_swin

Swin Transformer For Semantic Segmentation Of Human Body Images

This repo contains the supported code and configurations to reproduce Swin Transformer for semantic segmentation.

Goal

The goal is to do semantic segmentation on photos that can include one or more person.

Dataset source

The dataset is available on download here

Dataset preparation

In order to create annotation data I converted masks to label format [0,1].
The file structure of the dataset is shown below:

├── data
│   ├── full_body_tik_tok
│   │   ├── annotations
│   │   │   ├── training_1D
│   │   │   │   ├── xxx.png
│   │   │   │   ├── yyy.png
│   │   │   │   ├── zzz.png
│   │   │   ├── validation_1D
│   │   ├── images
│   │   │   ├── training
│   │   │   │   ├── xxx.png
│   │   │   │   ├── yyy.png
│   │   │   │   ├── zzz.png
│   │   │   ├── validation

Installation

See get started for installation

Train and test

Navigate to train and test in colab
Notes:

• Refer to config to see example
• Refer to custom dataset preparation
• Refer to custom model preparation

Web application

Navigate to flask app to make prediction in web
Note:

• Do not forget put your ngrok token

Results

Refer to youtube.

Name	mAcc	mIoU	Model
swin_tiny_patch4_window7_224	0.9291	0.7851	model
swin_base_patch4_window7_224	0.9015	0.7319	model
fcn_r101_d8	0.9798	0.9009	model
pspnet_r50-d8	0.9611	0.9356	model