segmentation-selectstar

Introduction

This is a prototype model of Pedestrian zone detection for blind people.

Separates sidewalk and driveway areas using Semantic Segmentation.

Sample Results	-

Original repository: https://github.com/jfzhang95/pytorch-deeplab-xception

Modified to run NIA SurfaceMasking dataset by [email protected]

Model

DeepLab v3+

[Paper] https://arxiv.org/abs/1802.02611

Training Surface Masking Dataset

1. Download NIA Surface Masking dataset from AIhub. (Not yet published)

   The Original NIA Surface Masking dataset consists of the following classes:

   class@attribute	Meaning
   alley@crosswalk	이면도로 - 횡단보도
   alley@damaged	이면도로 -파손
   alley@normal	이면도로 - 속성값 없음
   alley@speed_bump	이면도로 - 과속방지턱
   bike_lane	자전거도로
   braille_guide_blocks@damaged	점자블록 -파손
   braille_guide_blocks@normal	점자블록 -속성값 없음
   caution_zone@grating	주의구역 - 그레이팅
   caution_zone@manhole	주의구역 - 맨홀
   caution_zone@repair_zone	주의구역 - 보수구역
   caution_zone@stairs	주의구역 - 계단
   caution_zone@tree_zone	주의구역 - 가로수영역
   roadway@crosswalk	차도 - 횡단보도
   roadway@normal	차도 - 속성값없음
   sidewalk@asphalt	인도 - 아스팔트
   sidewalk@blocks	인도 - 보도블럭
   sidewalk@cement	인도 - 시멘트
   sidewalk@damaged	인도 - 파손
   sidewalk@other	인도 - 기타
   sidewalk@soil_stone	인도 - 흙,돌,비포장
   sidewalk@urethane	인도 - 우레탄

   But there are too many classes to do segmentation, so I reduced into 6 classes:

   New Class	Label	RGB Color
   background	0	[0, 0, 0]
   bike_lane	1	[255, 128, 0]
   caution_zone	2	[255, 0, 0]
   crosswalk	3	[255, 0, 255]
   guide_block	4	[255, 255, 0]
   roadway	5	[0, 0, 255]
   sidewalk	6	[0, 255, 0]

   Check settings.py for detailed classes info.

2. Generate mask images by running:

   1. modules/utils/surface_dataset_tools/surface_polygon.py
   2. modules/utils/surface_dataset_tools/split_dataset.py

   image	mask

3. Dataset structure should be like this.

   surface6
   ├── annotations
   │   ├── *.xml
   ├── images
   │   ├── *.jpg
   ├── masks
   │   ├── *.png
   ├── train.txt
   └── valid.txt
   

4. Install python packages

   Install Anaconda3 [https://www.anaconda.com/distribution/]
   conda create ml
   conda activate ml
   conda install conda
   conda install pytorch torchvision cudatoolkit=10.1 -c pytorch
   pip install tensorboardx, matplotlib
   

5. Edit training options settings.py

   Designate dataset directory
   ...
   elif dataset == 'surface':
       root_dir = '/home/super/Projects/dataset/surface6'
   ...
   

6. Run train.py

   1. On Windows: python train.py
   2. On Linux: python3 train.py
   3. On multi-gpu: CUDA_VISIBLE_DEVICES=0,1,2,3 python3 train.py

Download Trained Weights

Model Weights

Download and put it into ./run/surface/deeplab/model_iou_77.pth.tar

(Just create directory or Edit settings.py -> checkpoint, predict.py -> MODEL_PATH)

Settings for Reproduction

Trained on TitanXP x 4

Predict

1. Prepare 'mp4 video' or 'jpg images' to predict. And put it into 'test' directory.
2. Prepare trained model like model_iou_77.pth.tar
3. Edit RUN OPTIONS on predict.py

   MODEL_PATH, MODE, DATA_PATH, OUTPUT_PATH
   

4. Run predict.py
5. Output result will be saved to OUTPUT_PATH

Evaluate

1. Prepare dataset and trained model file.
2. Check settings.py options.
3. Run evaluate.py

Performance

Result of 2000 random selected validation set.

(fwIoU: Frequency Weighted Intersection over Union)

Acc	Acc_class	mIoU	fwIoU
91.46%	84.74%	77.29%	84.34%

IoU of each class

Class	IoU
background	85.40%
bike_lane	64.78%
caution_zone	57.19%
crosswalk	80.21%
guide_block	81.34%
roadway	85.69%
sidewalk	86.45%