Awesome

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.

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:

Results:

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

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.