[CVPR2024] Traffic Scene Parsing through the TSP6K Dataset

The dataset and code in TSP6K dataset. Code is implemented using an open-source semantic segmentation toolbox, MMsegmentation.

Installation

Please follow the installation instructions in mmsegmentation. In our environment, we use the following versions of different packages.

mmsegmentation==0.20.2
mmcv-full=1.4.0

Install the mmseg lib first

git clone https://github.com/PengtaoJiang/TSP6K.git
cd TSP6K/
pip install -v -e .

If you want to evaluate the iIoU score, please install the cityscapesscript lib

cd mmseg/datasets/cityscapesscripts/
python setup.py build install

Dataset Preparation

Download the dataset from this link and put them into /data/TSP6K/.

data
├── TSP6K
│   ├── image
│   ├── label
│   ├── split

You can also download the COCO-style instance bounding box annotations from this link.

Training

Train SegNext with the proposed Detail Refining Decoder using the following command

bash tools/dist_train.sh \
configs/tsp6k/segnext_base_1024x1024_160k_tsp6k_msaspp_rrm_5tokens_12heads.py \
8 --auto-resume  

Evaluation

Results and models

We provide the pre-trained segmentation models above. You can download them and directly evaluate them by

bash tools/dist_test.sh \
    configs/tsp6k/segnext_base_1024x1024_160k_tsp6k_msaspp_rrm_5tokens_12heads.py \
    ./work_dirs/segnext_base_1024x1024_160k_tsp6k_msaspp_rrm_5tokens_12heads/latest.pth \
    8 --out ./work_dirs/segnext_base_1024x1024_160k_tsp6k_msaspp_rrm_5tokens_12heads/results.pkl \
    --aug-test --eval mIoU  

Evaluate the segmentation model using the iIoU metric by

bash tools/dist_test.sh \
    configs/tsp6k/segnext_base_1024x1024_160k_tsp6k_msaspp_rrm_5tokens_12heads.py \
    ./work_dirs/segnext_base_1024x1024_160k_tsp6k_msaspp_rrm_5tokens_12heads/latest.pth \
    8 --out ./work_dirs/segnext_base_1024x1024_160k_tsp6k_msaspp_rrm_5tokens_12heads/results.pkl \
    --aug-test --eval cityscapes  

Citation

If you find the proposed TSP6K dataset and segmentation network are useful for your research, please cite

@inproceedings{jiang2024traffic,
  title={Traffic Scene Parsing through the TSP6K Dataset},
  author={Jiang, Peng-Tao and Yang, Yuqi and Cao, Yang and Hou, Qibin and Cheng, Ming-Ming and Shen, Chunhua},
  booktitle={Proceedings of the IEEE/CVF conference on computer vision and pattern recognition},
  year={2024}
}