liuguoyou / location-aware-upsampling-for-semantic-segmentation

The Pytorch implementation of "Location-aware Upsampling for Semantic Segmentation" (LaU)

Location-aware-Upsampling [arXiv]

Pytorch implementation of "Location-aware Upsampling for Semantic Segmentation" (LaU). If you are only interested in the upsampling part, please refer to LaU.md.

1. Dependencies :

• Python 3.5.6
• PyTorch 1.0.0
• GCC 7.3.0

2. Usage :

2.1 install requirements

conda create -n LaU python=3.5
conda activate LaU
pip install ninja cython numpy nose tqdm cffi==1.0.0 Pillow scipy requests torch==1.0.0 torchvision==0.2.2.post2 
cd ./LaU-reg
# install detail-api
cd ./detail-api/PythonAPI/
python setup.py install
# install LaU
cd ../../
python setup.py develop # python setup.py install
cp -r ../LaU ./encoding/models/
cd ./encoding/models/LaU
bash make.sh

2.2 Prepare data

cd LaU-reg
# E.g., download pcontext
python scripts/prepare_pcontext.py
# python scripts/prepare_ade20k.py

2.3 Train / Val / Test

cd LaU-reg/experiments/segmentation
# E.g., train and evaluate EncNet101-LaU4x-reg on pcontext
# train
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python train.py --dataset pcontext \
    --model encnet --dilated --lateral --aux --se-loss --offset-loss --batch-size 16\
    --backbone resnet101 --checkname encnet_resnet101_pcontext --no-val --offset-weight 0.3 --location-weight 0.1 --up-factor 4 --batch-size-per-gpu 2 --bottleneck-channel 64 --offset-branch-input-channel 512 --category 150 --base-size 520 --crop-size 480 --downsampled-input-size 60
# val [single]
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python test.py --dataset pcontext \
    --model encnet --dilated --lateral --aux --se-loss --offset-loss \
    --backbone resnet101 --resume /home/shuan/LaU-reg/experiments/segmentation/runs/encnet_resnet101_pcontext/encnet/encnet_resnet101_pcontext/179_checkpoint.pth.tar --split val --mode test --up-factor 4 --batch-size-per-gpu 1 --bottleneck-channel 64 --offset-branch-input-channel 512 --category 150 --base-size 520 --crop-size 480 --downsampled-input-size 60 # --ms  

Or

cd LaU-reg/experiments/segmentation
# E.g., train and evaluate EncNet101-LaU4x-reg on pcontext
bash ./scripts/encnet_res101_pcontext.sh 2>&1 | tee ./log.txt

More scripts can be found in LaU-reg/experiments/segmentation/scripts/.

3. Main difference between LaU and EncNet

1. We replace the original SegmentationLosses with OffsetLosses (defined in LaU-reg/encoding/nn/customize.y and used in LaU-reg/experiments/segmentation/train.py).
2. We add an LaU to the top of the decoder in EncNet (defined in LaU-reg/encoding/models/encnet.py).
3. The LaU module can be found in LaU-reg/encoding/models/util.py and LaU-reg/encoding/models/lau.py.

4. Acknowledgement :

We would like to thank FastFCN and Encoding for sharing their codes!

5. Cite :

If you find this project helpful, please cite

@misc{he2019locationaware,
    title={Location-aware Upsampling for Semantic Segmentation},
    author={Xiangyu He and Zitao Mo and Qiang Chen and Anda Cheng and Peisong Wang and Jian Cheng},
    year={2019},
    eprint={1911.05250},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}