Skip to content

Latest commit

 

History

History
53 lines (32 loc) · 4.62 KB

File metadata and controls

53 lines (32 loc) · 4.62 KB

Dataset Preparation

We curated training and testing data for each dataset separately and stored them as PKL files to load for training.

In the data folder, we sampled a small subset of data points to demonstrate the curated data structure and the usage of the training code. For demonstration with the sampled small-sized data, the batch size is set to 1 in the training configuration of each specific dataset — for instance, in ../working_para/working_dir_root_train_pascal_p.py. Please change the batch size for actual experiments.

The steps for preparing the full large-scale data for each dataset are listed below.

Abdominal Surgery Dataset

Download link with preprocessing code and instructions in package Download.

Following the instructions, generate PKL files for training and testing, and place them in /data/MICCAI/video_clips_pkl/ and /data/MICCAI_selected_GT/pkl/.

Thoracic Surgery Dataset

Please fill in this data request form Download preprocessing code and instructions in data package Download.

Following the instructions, generate PKL files for training and testing, and place them in /data/Thoracic/pkl/ and /data/Thoracic/annotated/pkl/.

Cholecystectomy Surgery Dataset

Download link with preprocessing code and instructions in package Download.

Following the instructions, generate PKL files for training and testing, and place them in /data/cholec80/output_pkl_croped/ and /data/cholecseg8k_working/output_pkl_croped/.

PASCAL VOC 2012

Please download the processed dataset from the PASCAL official data host.

Unzip the downloaded tgz file. We do not need the folders with saliency in their names. Please keep only the images/, SegmentationClass/, SegmentationClassAug/, and sets/ folders.

Finally, we also need the instance segmentation masks for evaluation.
Please download this file, unzip it, and take the VOCdevkit/VOC2012/SegmentationObject/ folder.

Then run the data pre-curation code data_pascal.py to generate PKLs for training and testing, and place them in /data/PASCAL/VOCtrainval_11-May-2012/pkl/train and /data/PASCAL/VOCtrainval_11-May-2012/pkl/val.

MS COCO 2017

Please download the data from their website.
Specifically, we need 2017 Train images [118K/18GB], 2017 Val images [5K/1GB], and 2017 Train/Val annotations [241MB].

Unzip them and you will get two image folders, train2017/ and val2017/, and two annotation files, instances_train2017.json and instances_val2017.json.
Please put the image folders under ./data/COCO/images/, and the annotation JSON files under ./data/COCO/annotations/.

Then run the data pre-curation code data_coco.py to generate PKLs for training and testing, and place them in /data/COCO/pkl/train and /data/COCO/pkl/val.

YTVIS Data

Download the instance segmentation (2019) version from the YouTube Object Segmentation Challenge official website Download.

Then run the data pre-curation code data_ytvis.py and data_ytvis_test.py separately to generate PKLs for training and testing, and place them in /data/YTVOS/pkl/ and /data/YTVOS/pkl_test/.

YTOBJ Data

Please download the YTOBJ dataset from the official website, which contains videos for 10 objects and annotations in .mat format (https://calvin-vision.net/datasets/youtube-objects-dataset/).

Unzip the files and use data_ytobj.py and data_ytobj_box_test.py separately to generate PKLs for training and testing, and place them in /data/YTOBJ/pkl/ and /data/YTOBJ/pkl_test/.