Foodseg-uecfoodpix

This repo implements the deeplabv3+ training for UECFoodPIX complete dataset. And this repository implements the baseline for FoodSAM: Any Food Segmentation.

Installation

a. Create a conda virtual environment and activate it.

conda create -n foodseg-uec python=3.8 -y
conda activate foodseg-uec

b. Install PyTorch and torchvision following the official instructions. Here we use PyTorch 1.10.1 and CUDA 11.3. You may also switch to other version by specifying the version number.

conda install pytorch==1.10.1 torchvision==0.12.2 cudatoolkit=11.3 -c pytorch -c conda-forge -y

c. Install MMCV following the official instructions.

pip install mmcv-full -f https://download.openmmlab.com/mmcv/dist/{cu_version}/{torch_version}/index.html

d. Clone this repo.

git clone https://github.com/HitBadTrap/Foodseg-uecfoodpix.git
cd Foodseg-uecfoodpix
pip install -e .  # or "python setup.py develop"

Testing

Run the following commands to evaluate the given checkpoint:

python tools/test.py [config] [checkpoint] --show-dir [output_dir] --show(optional)

You can append --show to generate visualization results in the output_dir/vis_image.

For our testing example, move the downloaded checkpoint file into ckpts directory, then run

python tools/test.py ./configs/deeplabv3plus/deeplabv3plus_r101-d8_4xb4-80k_uecfoodpix-320x320.py ./ckpts/best_mIoU_iter_24000.pth --show-dir output --show

Training

1. For single-gpu training, run the following command:

python tools/train.py [config]

2. For multi-gpu training, run the following commands:

bash tools/dist_train.sh [config] [num_gpu]

The default config is ./configs/deeplabv3plus/deeplabv3plus_r101-d8_4xb4-80k_uecfoodpix-320x320.py

For our training example:

# single-gpu training
python tools/train.py ./configs/deeplabv3plus/deeplabv3plus_r101-d8_4xb4-80k_uecfoodpix-320x320.py

# multi-gpu training
bash tools/dist_train.sh ./configs/deeplabv3plus/deeplabv3plus_r101-d8_4xb4-80k_uecfoodpix-320x320.py 2

Results

Method	mIou	aAcc	mAcc	Model	Training Log
deeplabV3+ (baseline)	65.61	88.20	77.56	Link	Link
FoodSAM	66.14	88.47	78.01

Acknowledgements

A large part of the code is borrowed from mmsegmentation

License

The model is licensed under the Apache 2.0 license.

Citation

If you want to cite our work, please use this:

@misc{lan2023foodsam,
      title={FoodSAM: Any Food Segmentation}, 
      author={Xing Lan and Jiayi Lyu and Hanyu Jiang and Kun Dong and Zehai Niu and Yi Zhang and Jian Xue},
      year={2023},
      eprint={2308.05938},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}