Automatic damage detection using anchor-free method and unmanned surface vessel, Automation in Construction, 133 (2022) 104017, by Zhili He, Shang Jiang, Jian Zhang and Gang Wu. Our paper is available.
-
Requirements
- python >= 3.5
- pytorch >= 1.1.0
- CUDA 10.0 and CUDNN 7.4
- Other requirements can be found in the requirements.txt.
-
Clone the repo
git clone https://github.com/hzlbbfrog/CenWholeNet.git
cd CenWholeNet
Or, you can "Download ZIP".
-
Compile DCN and nms You can refer to CenterNet to compile DCN and nms in advance.
By the way, CenWholeNet follows CenterNet directly. It is easy to write out CenWholeNet based on CenterNet. (facepalm) -
Rewrite Damage.py
For your own data set, you should rewrite Damage.py. -
Train your own data set
- Resnet
python train.py --log_name Resnet18 --dataset Damage --arch resnet --lr 5e-4 --lr_step 90,120 --batch_size 2 --num_epochs 60 --num_workers 2
- PAM
python train_seed.py --log_name Resnet18_PAM --dataset Damage --arch resnet_PAM --lr 5e-4 --lr_step 90,120 --batch_size 2 --num_epochs 60 --num_workers 2
- Test your own data set
- Resnet
python test.py --log_name Resnet18 --arch resnet
- PAM
python test.py --log_name Resnet18_PAM --arch resnet_PAM
- About the demo data
A reviewer ever mentioned that the images in the paper all include only one kind of defects. We synthesized some images with multiple types of defects. The results corroborate that the proposed methods have certain generalization ability. We want to extend our gratitude to the anonymous reviewer.
The official repo was not adopted because of some reasons.
A simple pytorch implementation version was used, which is simpler and easier to read.
What's more, a detailed compilation process is introduced in that repo. O(∩_∩)O
If you want to use CenterNet quickly, try it!
Faster R-CNN was compared in our paper.
You can access this repo to get the corresponding codes.
To tell the truth, it may be a little complicated to compile Faster R-CNN in Win 10. (facepalm)
YOLOv5 was also compared in our paper.
You can access this repo to get the corresponding codes.
You are very welcomed to cite our paper!
Because I have been busy these days, code is not optimized very well and some notes may be in Chinese.
I am really sorry for that.
However, if you have any questions, please do not hesitate to contact me!