Exploring Compact and Efficient Neural Networks for Real-Time Semantic Segmentation

Introduction

This is the official repository for CDDRNet. The code is modified based on HRNet-Semantic-Segmentation and DDRNet.pytorch.

Reference Environment

2080Ti*2
pytorch 1.8.1
see requirements

Usage

0. Dataset

(Cityscapes) Download the Cityscapes dataset, including gtFine (gtFine_trainvaltest.zip (241MB) + leftImg8bit_trainvaltest.zip (11GB)) and extra(leftImg8bit_trainextra.zip (44GB) + label generated from HRNet-Semantic-Segmentation Download Link)
(CamVid) Download the CamVid dataset. including original images/labels and extra images (gtFine of Cityscapes) + labels Download Link. Notice that the Camvid labels are processed for accelerating the training speed. An example of converting the original label to process label is utils/generate_camvid_train_labels.py.
You can also train the network without extra data. In this case, you will get a lower MIoU. (also modify the TRAIN_SET path in experiments/cityscapes and set to train_no_extra.lst in data/list/camvid and data/list/cityscapes)
If you got errors while training, check image paths listed in data/list first.

1. Train

python -m torch.distributed.launch --nproc_per_node=2 tools/train.py --cfg experiments/camvid/cddrnet.yaml
or
python -m torch.distributed.launch --nproc_per_node=2 tools/train.py --cfg experiments/cityscapes/cddrnet.yaml

2. Eval

python tools/eval.py --cfg experiments/camvid/cddrnet.yaml
python tools/eval.py --cfg experiments/cityscapes/cddrnet.yaml

3. Speed Test

cd utils/speed_test/
python cddrnet_eval_speed.py

Troubleshooting

(Train in a Server)Import cv2 error: libGL.so.1: cannot open shared object file: No such file or directory

pip install opencv-python-headless (should have the same version with opencv-python)

himlen1990 / cddrnet