Deep Dual-resolution Networks for Real-time and Accurate Semantic Segmentation of Road Scenes

Introduction

This is the unofficial code of Deep Dual-resolution Networks for Real-time and Accurate Semantic Segmentation of Road Scenes. the origin official is the official repository, and I borrowed most of the code from DDRNet.Pytorchthanks for their work.

<figure> <text-align: center;> <center> <img src="./images/ddrnet.png" alt="hrnet" title="" width="400" height="400" /> </center> </figcaption> </figure>

Main Change

1. Input 512*512;
2. Change DAPPM module ;
3. Data augmentation: random_brightness,random_RotateAndCrop, random_hue, random_saturation, random_contrast ...
4. Train on face segmentation datasetSemantic_Human_Matting

Quick start

1. Data preparation

You need to download the Semantic_Human_Mattingdatasets. and rename the folder face, then put the data under data folder.

└── data
  ├── face
       |————train_images
       |————train_labels
       |————val_images
       |————val_labels
  └── list

2. Pretrained model

download the pretrained model on imagenet or the segmentation model from the official，and put the files in ${PROJECT}/pretrained_models folder

3. TRAIN

download the imagenet pretrained model, and then train the model with 2 nvidia-3080

python tools/train_single.py --cfg experiments/face/ddrnet23_slim.yaml

Results

<figure> <text-align: center;> <center> <img src="./images/a242.jpg" alt="hrnet" title="" width="400" height="200" /> </center> </figcaption> </figure>

Train Custom Data

The only change is to write your own dataset, you can reference to ‘./lib/datasets’

Mobile Seg

follow TorchMobile,test with S855+ and take about 150 ms per image.

<figure> <text-align: center;> <center> <img src="./images/mobile.jpg" alt="hrnet" title="" width="400" height="500" /> </center> </figcaption> </figure>

TensorRT

https://github.com/midasklr/DDRNet.TensorRT

Test on RTX2070

| model | input | FPS | | -------------- | --------------- | ---- | | Pytorch-aug | （3,1024,1024） | 107 | | Pytorch-no-aug | （3,1024,1024） | 108 | | TensorRT-FP32 | （3,1024,1024） | 117 | | TensorRT-FP16 | （3,1024,1024） | 215 | | TensorRT-FP16 | （3,512,512） | 334 |

Pytorch-aug means augment=True.

Reference

[1] DDRNet

[2] the official repository