Skip to content

R3ab/ttpla_dataset

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

57 Commits
 
 
 
 
 
 
 
 
 
 

Repository files navigation

TTPLA: An Aerial-Image Dataset for Detection and Segmentation of Transmission Towers and Power Lines

TTPLA is a public dataset which is a collection of aerial images on Transmission Towers (TTs) and Powers Lines (PLs). This is the official repository of paper TTPLA: An Aerial-Image Dataset for Detection and Segmentation of Transmission Towers and Power Lines.

Screenshot

The repository includes:

  • The original images of TTPLA dataset with pixel level annotation in COCO format. The dataset images here (updated March 2021).
  • Splitting text files contain a list of images names after splitting to train.txt, validate.txt, and test.txt.
  • Weights of training models based on two different backbones and three different image sizes.

Preparation data:

  1. Modify resize_image_and_annotation-final.py to use the target image dimension (line 10). Then, call the script using python resize_image_and_annotation-final.py -t <images_path>. It will produce new folder called sized_data.

  2. Then call remove_void.py to remove void label if you would like to remove it. python remove_void.py -t <sized_images_path>. It will produce new folder called newjsons, you may renamed to whatever is fit.

  3. Based on three lists of train.txt, test.txt, and val.txt, split_jsons.py is used to split the created newjsons to three folders train , val, and test to prepare this before get the COCO json file.You can use the following command. python split_jsons.py -t newjsons/. It will produce new folder called splitting_jsons, you may renamed to whatever is fit.

  4. Use labelme2coco_2.py to get the COCO_json that used by Yolact. python labelme2coco_2.py splitting_jsons/train_jsons/. This step is done for three folders train_jsons , val_jsons, and test_jsons.

Tips to use our files directly

  • Install yolact Yolact.
  • Rename yolact folder to yolact700. Based on different sizes, it can rename also to yolact550 or yolact640.
  • In setp 1 in Prepration data, rename the generated sized_data folder name to data_700x700 and upload in yolact700/data/data_700x700. Based on different sizes, data_550x550 and data_640x360 are the other named folders with different sizes.
  • Use the suitable configuration from next table according to image size and backbone. Rename the picked config file to config.py and insert in yolact700/data/.
  • The generated json from step 4 in Prepration data, rename to train_coco_700x700, 2_test_json700, 2_val_json700 and put them into yolact700/data/ if you would like to use our config file directly or you can use any name and modify the pathes into config file.

Train Model:

For train image for example with size 700x700,

python train.py --config=yolact_img700_val_config --batch_size=8 --resume=weights/yolact_img550_108_12253_interrupt.pth

For evaluation,

python eval.py --config=yolact_img550_secondtest_config --mask_proto_debug --trained_model=weights/weights_img550_resnet50/yolact_img550_400_30061_resnet50_sep7_2217.pth --fast_nms=false

Evaluation:

Image Size Backbone configs weights
640 x 360 Resnet50 config_img640_resnet50_aspect.py yolact_img640_secondval_399_30000_resnet50.pth
550 x 550 Resnet50 config_img550_resnet50.py yolact_img550_399_30000_resnet50.pth
700 x 700 Resnet50 config_img700_resnet50.py yolact_img700_399_30000_resnet50.pth
640 x 360 Resnet101 config_img640_resnet101_aspect.py yolact_img640_secondval_399_45100_resnet101.pth
550 x 550 Resnet101 config_img550_resnet101.py yolact_img550_399_45100_resnet101_b8.pth
700 x 700 Resnet101 config_img700_resnet101.py yolact_img700_399_45100_resnet101_b8.pth

Results:

Average Precision for Different Deep Learning Models on TTPLA is reported in the following table

results

Citation:

@inproceedings{abdelfattah2020ttpla,
  title={TTPLA: An Aerial-Image Dataset for Detection and Segmentation of Transmission Towers and Power Lines},
  author={Abdelfattah, Rabab and Wang, Xiaofeng and Wang, Song},
  booktitle={Proceedings of the Asian Conference on Computer Vision},
  year={2020}
}

Contact:

For questions about our paper or code, please contact Rabab Abdelfattah.

About

aerial images dataset on transmission towers and power lines

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages