Segmentation performance of LMFFNet

Crop Size*	Dataset	Pretrained	Train type	mIoU	Params	Speed
512,1024	Cityscapes	No	trainval	75.1	1.35	118.9
1024,1024	Cityscapes	No	trainval	76.1	1.35	-
360,480	CamVid	No	trainval	69.1	1.35	116.4
720,960	CamVid	No	trainval	72.0	1.35	120.8

* Represents the resolution of the input image cropping in the training phase. We found that when the randomly cropped image is 1024x1024 in the training phase, the network can perform better. If the input image of LMFFNet is randomly cropped to 1024x1024 resolution, 76.1% mIoU can be achieved on cityscapes.

Preparation

You need to download the Cityscapes and CamVid datasets and place the symbolic links or datasets of the Cityscapes and CamVid datasets in the dataset directory. Our file directory is consistent with DABNet (https://github.com/Reagan1311/DABNet).

dataset
  ├── camvid
  |    ├── train
  |    ├── test
  |    ├── val 
  |    ├── trainannot
  |    ├── testannot
  |    ├── valannot
  |    ├── camvid_trainval_list.txt
  |    ├── camvid_train_list.txt
  |    ├── camvid_test_list.txt
  |    └── camvid_val_list.txt
  ├── cityscapes
  |    ├── gtCoarse
  |    ├── gtFine
  |    ├── leftImg8bit
  |    ├── cityscapes_trainval_list.txt
  |    ├── cityscapes_train_list.txt
  |    ├── cityscapes_test_list.txt
  |    └── cityscapes_val_list.txt           

How to run

1 Training

1.1 Cityscapes

python train.py

1.2 CamVid

python python train.py --dataset camvid --train_type trainval --max_epochs 1000 --lr 1e-3 --batch_size 8

2 Testing

2.1 Cityscapes

python predict.py --dataset ${camvid, cityscapes} --checkpoint ${CHECKPOINT_FILE}

To convert the training lables to class lables.

python trainID2labelID.py Package the file into xxx.zip Submit the zip file to https://www.cityscapes-dataset.com/submit/. You can get the results from the https://www.cityscapes-dataset.com/submit/.

2.2 CamVid

python test.py --dataset camvid --checkpoint ${CHECKPOINT_FILE}

4. fps

python eval_forward_time.py --size 512,1024

To be continue...

Citation

@article{shi2022lmffnet,
 title={LMFFNet: A Well-Balanced Lightweight Network for Fast and Accurate Semantic Segmentation},
 author={Shi, Min and Shen, Jialin and Yi, Qingming and Weng, Jian and Huang, Zunkai and Luo, Aiwen and Zhou, Yicong},
 journal={IEEE Transactions on Neural Networks and Learning Systems},
 year={2022},
 publisher={IEEE}
}

Reference

https://github.com/xiaoyufenfei/Efficient-Segmentation-Networks

https://github.com/Reagan1311/DABNet