Skip to content

Latest commit

 

History

History
93 lines (73 loc) · 3.65 KB

README.md

File metadata and controls

93 lines (73 loc) · 3.65 KB
Raw

MVSS-Baseline

This repository provides the baseline code of MVSS task, i.e., MVNet.

Installation

The code requires python>=3.7, as well as pytorch>=1.9 and torchvision>=0.10. Please follow the instructions here to install both PyTorch and TorchVision dependencies. Installing both PyTorch and TorchVision with CUDA support is strongly recommended. Then, please clone this repo.

  1. Clone this repo.

    $ git clone https://github.com/jiwei0921/MVSS-Baseline.git
$ cd MVSS-Baseline-main/MVNet

  2. Install dependencies.

    $ conda create -n MVSS
$ conda activate MVSS
$ conda install pytorch torchvision torchaudio cudatoolkit=11.3 -c pytorch
$ pip install scipy
$ pip install tqdm
$ pip install visdom
$ pip install matplotlib

Getting Started

First download the MVSeg dataset and pretrained model ckpt. Then the model can be used in just a few adaptions to start training:

  1. Set your MVSeg dataset path in datasets/mvss_dataset.py and put the ckpt in file save/*.
  2. Perform training from scratch, with bash train.sh in two stage. First for warm-up, second for MVNet.
    or Perform training based on our warm-up ckpt, with bash train.sh for second stage.

Meanwhile, the segmentation maps can be generated by loading the pretrained model ckpt, with:

  1. Set your MVSeg dataset path in datasets/mvss_dataset.py and put the ckpt in file save/*.
  2. Specify testset name in test.sh, e.g., --split-mode test, or --split-mode test_night.
  3. Perform inference, with bash test.sh

Dataset ColorMap

Here is the reference for MVSeg dataset color visualization.

[
        (0, 0, 0),          # 0:    background(unlabeled)
        (0, 0, 142),        # 1:    Car
        (0, 60, 100),       # 2:    Bus
        (0, 0, 230),        # 3:    Motorcycle
        (119, 11, 32),      # 4:    Bicycle
        (255, 0, 0),        # 5:    Pedestrian
        (0, 139, 139),      # 6:    Motorcyclist
        (255, 165, 150),    # 7:    Bicyclist
        (192, 64, 0),       # 8:    Cart
        (211, 211, 211),    # 9:    Bench
        (100, 33, 128),     # 10:   Umbrella
        (117, 79, 86),      # 11:   Box
        (153, 153, 153),    # 12:   Pole
        (190, 122, 222),    # 13:   Street_lamp
        (250, 170, 30),     # 14:   Traffic_light
        (220, 220, 0),      # 15:   Traffic_sign
        (222, 142, 35),     # 16:   Car_stop
        (205, 155, 155),    # 17:   Color_cone
        (70, 130, 180),     # 18:   Sky
        (128, 64, 128),     # 19:   Road
        (244, 35, 232),     # 20:   Sidewalk
        (0, 0, 70),         # 21:   Curb
        (107, 142, 35),     # 22:   Vegetation
        (152, 251, 152),    # 23:   Terrain
        (70, 70, 70),       # 24:   Building
        (110, 80, 100)      # 25:   Ground
]

Citation

@InProceedings{ji2023mvss,
      title     = {Multispectral Video Semantic Segmentation: A Benchmark Dataset and Baseline},
      author    = {Ji, Wei and Li, Jingjing and Bian, Cheng and Zhou, Zongwei and Zhao, Jiaying and Yuille, Alan L. and Cheng, Li},
      booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
      month     = {June},
      year      = {2023},
      pages     = {1094-1104}
}

Acknowledgement

This repository was originally built from LMANet. It was modified and extended to support our multispectral video setting.