❗ This is cloned repository!

This repository is cloned from chenjun2hao/DDRNet.pytorch and modified for research

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. which achieve state-of-the-art trade-off between accuracy and speed on cityscapes and camvid, without using inference acceleration and extra data!on single 2080Ti GPU, DDRNet-23-slim yields 77.4% mIoU at 109 FPS on Cityscapes test set and 74.4% mIoU at 230 FPS on CamVid test set.

The code mainly borrows from HRNet-Semantic-Segmentation OCR and the official repository, thanks for their work.

^{A comparison of speed-accuracy trade-off on Cityscapes test set.}

Requirements

torch>=1.7.0
cudatoolkit>=10.2

Cityscapes Data Preparation

Download two files below from Cityscapes.to the \${CITYSCAPES_ROOT}
- leftImg8bit_trainvaltest.zip
- gtFine_trainvaltest.zip
Unzip them

Rename the folders like below

└── cityscapes
 ├── leftImg8bit
     ├── test
     ├── train
     └── val
 └── gtFine
     ├── test
     ├── train
     └── val

Update some properties in {REPO_ROOT}/experiments/cityscapes/${MODEL_YAML} like below

DATASET:
 DATASET: cityscapes
 ROOT: ${CITYSCAPES_ROOT}
 TEST_SET: 'cityscapes/list/test.lst'
 TRAIN_SET: 'cityscapes/list/train.lst'
 ...

Pretrained Models

The official repository provides pretrained models for Cityscapes

Download the pretrained model to \${MODEL_DIR}

Update MODEL.PRETRAINED and TEST.MODEL_FILE in {REPO_ROOT}/experiments/cityscapes/${MODEL_YAML} like below

...
MODEL:
 ...
 PRETRAINED: "${MODEL_DIR}/${MODEL_NAME}.pth"
 ALIGN_CORNERS: false
 ...
TEST:
 ...
 MODEL_FILE: "${MODEL_DIR}/${MODEL_NAME}.pth"
 ...

Validation

Execute the command below to evaluate the model on Cityscapes-val

cd ${REPO_ROOT}
python tools/eval.py --cfg experiments/cityscapes/ddrnet23_slim.yaml

model	OHEM	Multi-scale	Flip	mIoU	FPS	E2E Latency (s)	Link
DDRNet23_slim	Yes	No	No	77.83	91.31	0.062	official
DDRNet23_slim	Yes	No	Yes	78.42	TBD	TBD	official
DDRNet23	Yes	No	No	79.51	TBD	TBD	official
DDRNet23	Yes	No	Yes	79.98	TBD	TBD	official

mIoU denotes an mIoU on Cityscapes validation set.

FPS is measured by following the test code provided by SwiftNet. (Refer to speed_test from lib/utils/utils.py for more details.)

E2E Latency denotes an end-to-end latency including pre/post-processing.

FPS and latency are measured with batch size 1 on RTX 2080Ti GPU and Threadripper 2950X CPU.

Note

with the ALIGN_CORNERS: false in ***.yaml will reach higher accuracy.

TRAIN

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

cd ${PROJECT}
python -m torch.distributed.launch --nproc_per_node=2 tools/train.py --cfg experiments/cityscapes/ddrnet23_slim.yaml

the own trained model coming soon

OWN model

model	Train Set	Test Set	OHEM	Multi-scale	Flip	mIoU	Link
DDRNet23_slim	train	eval	Yes	No	Yes	77.77	Baidu/password:it2s
DDRNet23_slim	train	eval	Yes	Yes	Yes	79.57	Baidu/password:it2s
DDRNet23	train	eval	Yes	No	Yes	~	None
DDRNet39	train	eval	Yes	No	Yes	~	None

Note

set the ALIGN_CORNERS: true in ***.yaml, because i use the default setting in HRNet-Semantic-Segmentation OCR.
Multi-scale with scales: 0.5,0.75,1.0,1.25,1.5,1.75. it runs too slow.
from ydhongHIT, can change the align_corners=True with better performance, the default option is False

Reference

[1] HRNet-Semantic-Segmentation OCR branch

[2] the official repository

swoook / ddrnet

readme