数据集获取

git clone https://github.com/SURA23/Sinus-Surgery-Endoscopic-Image-Datasets.git

curl -L -o repo.zip https://github.com/SURA23/Sinus-Surgery-Endoscopic-Image-Datasets/archive/refs/heads/master.zip
unzip repo.zip

源站 README 摘录（使用方式）

Sinus-Surgery-Endoscopic-Image-Datasets

Novel image segmentation datasets collected from endoscopic videos of sinus surgery processes
Examples
UW Sinus Surgery Cadaver/Live (UW-Sinus-Surgery-C/L) Dataset:
Link: https://digital.lib.washington.edu/researchworks/handle/1773/45396
This dataset was developed at the University of Washington’'s BioRobotics Lab (http://brl.ee.washington.edu). It has endoscopic sinus surgery images with manual annotations for surgical instrument segmentation task. The dataset was collected from endoscopic sinus surgeries performed by surgeons, which are featured by dexterous tip motion, narrow operation space and close lens-object distance. More details can be found in the references [1,2].

UW-Sinus-Surgery-C/L consists of two sub-datasets: cadaver surgery dataset (UW-Sinus-Surgery-C) and live surgery dataset (UW-Sinus-Surgery-L). In each dataset, the folder “images” has endoscopic images and the folder “labels” has the segmentation ground truths (0 stands for background and 1 for surgical instrument)

The dataset folder is arranged as follows:

|- readme.txt

|- cadaver

  |-images

  |-labels

|

|—— live

  |-images

  |-labels

Image naming:

i) Sinus-Surgery-C dataset: S[video_ID]_[frame_index]

ii) Sinus-Surgery-L dataset: L[video_ID]_[frame_index]

[video_ID] shows from which video the image was extracted, [frame_index] is the frame’'s index in the corresponding video.

3-fold cross-validation experimental setup:

In [2], the image segmentation performances were evaluated based on K-fold cross-validation method, to avoid the bias caused by dataset split. The 3-fold crossvalidation is conducted with the setup below:

i) Sinus-Surgery-C dataset: fold 1: 1st~4th videos; fold 2: 5th-7th videos; fold 3: 8th-10th videos

ii) Sinus-Surgery-L dataset: fold 1: 1st video; fold 2: 2nd video; fold 3: 3rd video

If you are trying to follow the experiments in [2], please contact {fbqin ‘‘at’’ uw.edu} or {qinfangbo2013 ‘‘at’’ ia.ac.cn} to request the datasets which had already been organized to 3-fold.

Contact

For any questions and suggestions, please contact us via shl102@eng.ucsd.edu and qinfangbo2013@ia.ac.cn.

If you find this dataset helpful in your research, please cite our papers [1] and [2].

References

[1] S. Lin, F. Qin, et al., “LC-GAN: Image-to-image translation based

数据加载示例（图像类）

from PIL import Image
import glob, os

files = (glob.glob(os.path.join(path, "**", "*.png"), recursive=True)
       + glob.glob(os.path.join(path, "**", "*.jpg"), recursive=True)
       + glob.glob(os.path.join(path, "**", "*.tif"), recursive=True))
print("图像文件数:", len(files))
img = Image.open(files[0]); print("尺寸/模式:", img.size, img.mode)

# torchvision Dataset 方式:
# from torchvision import datasets
# ds = datasets.ImageFolder(path)  # 要求 子目录=类别

目录组织与标注格式以源站说明和下载后实际文件为准。

完整仓库：github.com/SURA23/Sinus-Surgery-Endoscopic-Image-Datasets