Road Segmentation

• Development environment

  ubuntu 18.04, tensorflow 2.0.0, opencv-python 4.2.0.32, numpy 1.18.2

• model

  FCN(test x)

  UNet (Test complete)

  PSPNet(Modification Required)

  PSPUNet(PSPNet + Unet, Test complete)

  ICNet(Modification Required)

• Learning environment

  • Dataset - AI Hub sidewalk walking image

  • train data : 38000, val data : 12800

  • IMG_WIDTH = 480

  • IMG_HEIGHT = 272

  • n_classes = 7

  • data argumentation - random_flip_horizon, random brightness

  • learning rate : 1e-4 -> epoch>10 lr decay(1e-5)

• Class

class	label
Background	0
Bike_lane	1
Caution_zone	2
Crosswalk	3
braille_guide_blocks	4
Roadway	5
Sidewalk	6

detail

• Quick start

I provide a pretraining model weight. (pspunet loss 0.3160 mIoU 74.5% acc 90.2%)

git clone https://github.com/JunHyeok96/Road-Segmentation.git
cd road_segmentation
python3 demo.py #before run command, you have to set your test video path in demo.py 

PSPUnet

UNet

• performance evaluation

  TEST GPU - RTX2060 SUPER

model	accuracy	loss	mIoU	FPS	Size
PSPUnet	90.2%	0.3160	74.5%	24.8	39.6MB
UNet	89.1%	0.3570	70.9%	22.7	131MB