dataset.md

March 29, 2024 ยท View on GitHub

Data

Stage 1: Pre-training Dataset

  • Please download the transformed annotations of each dataset from Stage-1 Training Annotations.

  • Please download the image from the official source.

DataSourceDataSource
COCO 2014DownloadVOCdevkitDownload
COCO 2017DownloadDocBankDownload
Visual GenomeDownloadDocLayNetDownload
Object365DownloadPubLayNetDownload
OpenImageDownloadCurvedSynText150kDownload
V3DetDownloadICDAR2013Download
ADE20kDownloadMLT2017Download
CityscapesDownloadMLT2019Download
cocostuff 10kDownloadTotalTextDownload
cocostuff 164kDownloadAITWDownload

Important notice: Visual Genome should contain all the vg images(VG_100K and VG_100K_2). Merge the image data from the VG_100K and VG_100K_2 folders into one.

  • In each annotation JSON file, update the image path to reflect the location of the downloaded image data.

Stage 2: Fine-tuning Dataset

  • Please download the transformed annotations of each dataset from Stage-2 Training Annotations.

  • Please download the image from the official source. The data for stages beyond stage 1 is list below:

DataSourceDataSource
OpenPsgGCGDownloadSeeClickDownload
GRITDownloadMulti-PanelDownload
Flicker30KDownloadOsprey-724KDownload
M6DocDownloadLaionGPT4vDownload
VCRDownloadShareGPT4vDownload
  • In each annotation JSON file, update the image path to reflect the location of the downloaded image data.