Prepare Datasets

February 2, 2024 · View on GitHub

Expected dataset structure:

Dataset structure:

openvis
├── datasets
│   ├── ytvis_2019
│   │   ├── {train, valid, test}.json
│   │   ├── {train, valid, test}/
│   │   │   ├── JPEGImages/
│   │   │   ├── Annotations/
│   ├── ytvis_2021
│   │   ├── {train, valid, test}/
│   │   │   ├── JPEGImages/
│   │   │   ├── instances.json
│   ├── ovis
│   │   ├── annotations_{train, valid, test}.json/
│   │   ├── Images/
│   │   │   ├── {train, valid, test}/
│   ├── burst
│   │   ├── annotations/
│   │   │   ├── train/
│   │   │   │   ├── train.json
│   │   │   ├── {val, test}/
│   │   │   │   ├── all_classes.json
│   │   │   │   ├── common_classes.json
│   │   │   │   ├── uncommon_classes.json
│   │   ├── frames/
│   │   │   ├── {train, val, test}/
│   ├── lvvis
│   │   ├── {train, val}_instances.json
│   │   ├── {train, val}_ytvis_style.json
│   │   ├── {train, val}/
│   │   │   ├── JPEGImages/
│   ├── coco
│   │   ├── coco2ytvis2019_{train, val}.json
│   │   ├── coco2ytvis2021_{train, val}.json
│   │   ├── coco2ovis_{train, val}.json
│   │   ├── {train, val}2017/
│   │   ├── annotations/
│   │   │   ├── instances_{train, val}2017.json

Youtube-VIS

openvis
├── datasets
│   ├── ytvis_2019
│   │   ├── {train, valid, test}.json
│   │   ├── {train, valid, test}/
│   │   │   ├── JPEGImages/
│   │   │   ├── Annotations/
│   ├── ytvis_2021
│   │   ├── {train, valid, test}/
│   │   │   ├── JPEGImages/
│   │   │   ├── instances.json

youtube-vis 2019:

  1. register 2nd youtube vos challenge
  2. download {train,valid,test}.zip from youtube-vis 2019 frames google drive
  3. download {train,valid,test}.json from youtube-vis 2019 annotations google drive
  4. unzip {train,valid,test}.zip to datasets/ytvis_2019/{train,valid,test}/
  5. put {train,valid,test}.json to datasets/ytvis_2019/{train,valid,test}.json

youtube-vis 2021:

  1. register 3rd youtube vos challenge
  2. download {train,valid,test}.zip from youtube-vis 2021 frames google drive
  3. unzip {train,valid,test}.zip to datasets/ytvis_2021/{train,valid,test}/

OVIS

openvis
├── datasets
│   ├── ovis
│   │   ├── annotations_{train, valid, test}.json/
│   │   ├── Images/
│   │   │   ├── {train, valid, test}/
  1. register Occluded Video Instance Segmentation
  2. download {train,valid,test}.zip from OVIS frames google drive
  3. download {train,valid,test}.json from OVIS annotations google drive
  4. unzip {train,valid,test}.zip to datasets/ovis/Images/{train,valid,test}/
  5. put annotations/{train,valid,test}.json to datasets/ovis/annotations_{train,valid,test}.json

BURST

openvis
├── datasets
│   ├── burst
│   │   ├── annotations/
│   │   │   ├── train/
│   │   │   │   ├── train.json
│   │   │   ├── {val, test}/
│   │   │   │   ├── all_classes.json
│   │   │   │   ├── common_classes.json
│   │   │   │   ├── uncommon_classes.json
│   │   ├── frames/
│   │   │   ├── {train, val, test}/
  1. download videos (except AVA & HACS videos) from TAO data:
wget "https://motchallenge.net/data/1-TAO_TRAIN.zip" 
wget "https://motchallenge.net/data/2-TAO_VAL.zip" 
wget "https://motchallenge.net/data/3-TAO_TEST.zip"
unzip 1-TAO_TRAIN.zip
unzip 2-TAO_VAL.zip
unzip 3-TAO_TEST.zip
  1. put uncompressed folders to datasets/burst/frames/{train, val, test}
  2. sign in MOTchallenge and download AVA & HACS videos
  3. download annotations from BURST-benchmark: wget https://omnomnom.vision.rwth-aachen.de/data/BURST/annotations.zip
  4. uncompress annotations.zip to datasets/burst/annotations

LVVIS

openvis
├── datasets
│   ├── lvvis
│   │   ├── {train, val}_instances.json
│   │   ├── {train, val}_ytvis_style.json
│   │   ├── {train, val}/
│   │   │   ├── JPEGImages/
  1. download videos from LV-VIS github;
    • download train.zip from google drive;
    • download val.zip from google drive;
    • uncompressed {train, val}.zip to datasets/lvvis/{train, val}/JPEGImages/;
  2. download annotations from LV-VIS github;
    • download train_instances.json from google drive;
    • download val_instances.json from google drive;
    • put {train, val}_instances.json to datasets/lvvis/{train, val}_instances.json;
  3. convert annotations to youtube-vis style:
python datasets/lvvis2ytvis.py

COCO

$ROOT
├── datasets
│   ├── coco
│   │   ├── coco2ytvis2019_{train, val}.json
│   │   ├── coco2ytvis2021_{train, val}.json
│   │   ├── coco2ovis_{train, val}.json
│   │   ├── {train, val}2017/
│   │   ├── annotations/
│   │   │   ├── instances_{train, val}2017.json
  1. download images from COCO homepage;
    • download train2017.zip, val2017.zip;
    • uncompress {train, val}2017,zip to datasets/coco/{train, val}2017/;
  2. download annotations from COCO homepage;
    • download annotations_trainval2017.zip;
    • uncompress annotations_trainval2017.zip to datasets/coco/annotations; 3 convert annotations to youtube-vis style:
python datasets/coco2ytvis.py