helonin / circle-u-net Goto Github PK

This is implementation of Circle U Net for image segmentation

Circle-U-Net: An Efficient Architecture for Semantic Segmentation , published in Algorithms Journal 2021

Dataset : ICG - TUGRAZ dataset

Installations

tensorflow-gpu==2.1.4

---dataset
          |
          |---data--
                    |
                    |
                    |
                    |------icg_drone
                    |               |
                    |               |-----train_frames--
                    |               |                   |-----train------
                    |               |                   |                |001.jpg
                    |               |                   |                |002.jpg
                    |               |                   |                |003.jpg
                    |               |-----train_masks---
                    |               |                   |-----train------
                    |               |                   |                |001.jpg
                    |               |                   |                |002.jpg
                    |               |                   |                |003.jpg
                    |               |
                    |               |-----val_frames----
                    |               |                   |-----val--------
                    |               |                   |                |055.jpg
                    |               |                   |                |056.jpg
                    |               |                   |                |057.jpg
                    |               |-----val_masks-----
                    |               |                   |-----val--------
                    |               |                   |                |055.jpg
                    |               |                   |                |056.jpg
                    |               |                   |                |057.jpg
                    |
                    |---label_color.txt

In ICG semantic drone dataset ,

• "train_frames" could be taken from - semantic_drone_dataset_semantics_v1.1\semantic_drone_dataset\training_set\images

• "train_masks" could be taken from - semantic_drone_dataset_semantics_v1.1\semantic_drone_dataset\training_set\gt\semantic\label_images

Please randomly seperate train and val set as you like make sure there are 360 for training and 40 for testing

Train frames

all other images except in val set

Train masks

all other images except in val set

Val frames

3,19,53,71,89,104,122,139,182,177,216,225,244,263,290,304,320,332,367,386 412,421,438,476,489,507,524,545,567,583,584,585,586,587,588,590,591,592,593,593,594

Val masks

3,19,53,71,89,104,122,139,182,177,216,225,244,263,290,304,320,332,367,386 412,421,438,476,489,507,524,545,567,583,584,585,586,587,588,590,591,592,593,593,594

Attenation Unet

python drone_main.py -d "camvid_small" -idir "dataset/icg_drone/data/" -m "att_unet" -ht 256 -w 256 -bs 5 --loss tversky --num_epochs 60

Resnet101 Unet

python drone_main.py -d "camvid_small" -idir "dataset/icg_drone/data/" -m "res_unet" -ht 256 -w 256 -bs 5 --loss tversky --num_epochs 60

Unet

python drone_main.py -d "camvid_small" -idir "dataset/icg_drone/data/" -m "tiny_unet" -ht 256 -w 256 -bs 5 --loss tversky --num_epochs 60

Circlenet - Tversky loss

python drone_main.py -d "camvid_small" -idir "dataset/icg_drone/data/" -m "circlenet" -ht 256 -w 256 -bs 5 --loss tversky --num_epochs 60

Circlenet - Categorical cross entropy

python drone_main.py -d "camvid_small" -idir "dataset/icg_drone/data/" -m "circlenet" -ht 256 -w 256 -bs 5 --loss CCE --num_epochs 60

Circlenet with attention - Tversky loss

python drone_main.py -d "camvid_small" -idir "dataset/icg_drone/data/" -m "circle_att_101" -ht 256 -w 256 -bs 5 --loss tversky --num_epochs 60

Circlenet with attention - Categorical cross entropy

python drone_main.py -d "camvid_small" -idir "dataset/icg_drone/data/" -m "circle_att_101" -ht 256 -w 256 -bs 5 --loss CCE --num_epochs 60

Attention unet - Categorical cross entropy

python drone_main.py -d "camvid_small" -idir "dataset/icg_drone/data/" -m att_unet -ht 256 -w 256 -bs 5 --loss CCE --num_epochs 60

Resunet - Categorical cross entropy

python drone_main.py -d "camvid_small" -idir "dataset/icg_drone/data/" -m "res_unet" -ht 256 -w 256 -bs 5 --loss CCE --num_epochs 60

Squeezeunet -CCE

python drone_main.py -d "camvid_small" -idir "dataset/icg_drone/data/" -m new_squeezenet -ht 256 -w 256 -bs 5 --loss CCE --num_epochs 60

Evaluating model and predicting images

python evaluate.py -d "camvid" -idir "dataset/camvid/data/" -mt "squeeze_unet_keras" -m "camvid_model_5_epochs.h5" -ht 256 -w 256

• Growth of Unet
• Unet - root

Last two years SOTA papers

• UNet
• UNet++
• Att_UNet
• ResUNet
• RexUnet
• Adversarial Learning
• NAS-Unet

CVPR

• Eff-UNet
• Feedback U-net
• Enhanced rotation equivariant Unet

ICCV

• Reccurent Unets

• BCDU-Net - https://github.com/rezazad68/BCDU-Net

• GAN based_on UNet

• RUNet for super resolution

Please site our paper if you use this code in your own work:

Citations:

Sun, Feng, Ajith K. V, Guanci Yang, Ansi Zhang, and Yiyun Zhang. 2021. "Circle-U-Net: An Efficient Architecture for Semantic Segmentation" Algorithms 14, no. 6: 159. https://doi.org/10.3390/a14060159