Temporal Memory Attention for Video Semantic Segmentation, arxiv

Introduction

We propose a Temporal Memory Attention Network (TMANet) to adaptively integrate the long-range temporal relations over the video sequence based on the self-attention mechanism without exhaustive optical flow prediction. Our method achieves new state-of-the-art performances on two challenging video semantic segmentation datasets, particularly 80.3% mIoU on Cityscapes and 76.5% mIoU on CamVid with ResNet-50.

Updates

2021/1: TMANet training and evaluation code released.

Usage

• Install mmseg

  • Please refer to mmsegmentation to get installation guide.
  • This repository is based on mmseg-0.7.0 and pytorch 1.6.0.

• Clone the repository.

  git clone https://github.com/wanghao9610/TMANet.git
  cd TMANet
  pip install -e .

• Prepare the datasets

  • Download Cityscapes dataset and Camvid dataset.

  • For Camvid dataset, we need to extract frames from downloaded videos, please view the code on ./tools/convert_datasets/.

  • Put the converted datasets on ./data/camvid and ./data/cityscapes path.

    File structure of video semantic segmentation dataset is as followed.

    ├── data                                              ├── data                  
    │   ├── cityscapes                                    │   ├── camvid
    │   │   ├── gtFine                                    │   │   ├── images
    │   │   │   ├── train                                 │   │   │   ├── train
    │   │   │   │   ├── xxx{img_suffix}                   │   │   │   │   ├── xxx{img_suffix}
    │   │   │   │   ├── yyy{img_suffix}                   │   │   │   │   ├── yyy{img_suffix}
    │   │   │   │   ├── zzz{img_suffix}                   │   │   │   │   ├── zzz{img_suffix}
    │   │   │   ├── val                                   │   │   │   ├── val
    │   │   ├── leftImg8bit                               │   │   ├── annotations
    │   │   │   ├── train                                 │   │   │   ├── train
    │   │   │   │   ├── xxx{seg_map_suffix}               │   │   │   │   ├── xxx{seg_map_suffix}
    │   │   │   │   ├── yyy{seg_map_suffix}               │   │   │   │   ├── yyy{seg_map_suffix}
    │   │   │   │   ├── zzz{seg_map_suffix}               │   │   │   │   ├── zzz{seg_map_suffix}
    │   │   │   ├── val                                   │   │   │   ├── val
    │   │   ├── leftImg8bit_sequence                      │   │   ├── image_sequence
    │   │   │   ├── train                                 │   │   │   ├── train
    │   │   │   │   ├── xxx{sequence_suffix}              │   │   │   │   ├── xxx{sequence_suffix}
    │   │   │   │   ├── yyy{sequence_suffix}              │   │   │   │   ├── yyy{sequence_suffix}
    │   │   │   │   ├── zzz{sequence_suffix}              │   │   │   │   ├── zzz{sequence_suffix}
    │   │   │   ├── val                                   │   │   │   ├── val
    

• Evaluation

  • Download the trained models for Cityscapes and Camvid. And put them on ./work_dirs/{config_file}
  • Run the following command(on Cityscapes):

  sh eval.sh configs/video/cityscapes/tmanet_r50-d8_769x769_80k_cityscapes_video.py

• Training

  • Please download the pretrained ResNet-50 model, and put it on ./init_models .
  • Run the following command(on Cityscapes):

  sh train.sh configs/video/cityscapes/tmanet_r50-d8_769x769_80k_cityscapes_video.py

  Note: the above evaluation and training shell commands execute on Cityscapes, if you want to execute evaluation or training on Camvid, please replace the config file on the shell command with the config file of Camvid.

Citation

If you find TMANet is useful in your research, please consider citing:

@misc{wang2021temporal,
    title={Temporal Memory Attention for Video Semantic Segmentation}, 
    author={Hao Wang and Weining Wang and Jing Liu},
    year={2021},
    eprint={2102.08643},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}

Acknowledgement

Thanks mmsegmentation contribution to the community!