pytorch

STDC

Model Description

STDC (Short-Term Dense Concatenate network) is an efficient deep learning model for semantic segmentation that reduces structural redundancy while maintaining high performance. It gradually decreases feature map dimensions and aggregates them for image representation. The model incorporates a Detail Aggregation module in the decoder to enhance spatial information learning in low-level layers. By fusing both low-level and deep features, STDC achieves accurate segmentation results with optimized computational efficiency, making it particularly suitable for real-time applications.

Supported Environments

GPU	IXUCA SDK	Release
BI-V100	3.0.0	23.06

Model Preparation

Prepare Resources

Go to visit Cityscapes official website, then choose 'Download' to download the Cityscapes dataset.

Specify /path/to/cityscapes to your Cityscapes path in later training process, the unzipped dataset path structure should look like:

cityscapes/
├── gtFine
│   ├── test
│   ├── train
│   │   ├── aachen
│   │   └── bochum
│   └── val
│       ├── frankfurt
│       ├── lindau
│       └── munster
└── leftImg8bit
    ├── train
    │   ├── aachen
    │   └── bochum
    └── val
        ├── frankfurt
        ├── lindau
        └── munster

mkdir -p data/
ln -s /path/to/cityscapes data/

Install Dependencies

# Install libGL
## CentOS
yum install -y mesa-libGL
## Ubuntu
apt install -y libgl1-mesa-glx

# install mmsegmentation
git clone -b v1.2.2 https://github.com/open-mmlab/mmsegmentation.git --depth=1
cd mmsegmentation/
pip install -v -e .

pip install ftfy

Model Training

# Training on single card
python3 tools/train.py configs/stdc/stdc1_4xb12-80k_cityscapes-512x1024.py

# Training on mutil-cards
sed -i 's/python /python3 /g' tools/dist_train.sh
bash tools/dist_train.sh configs/stdc/stdc1_4xb12-80k_cityscapes-512x1024.py 8

Model Results

GPUs	Crop Size	Lr schd	FPS	mIoU
BI-V100 x8	512x1024	20000	39.38	70.74

References

• mmsegmentation