Skip to content

Data.md

Latest commit

 

History

History
49 lines (44 loc) · 1.63 KB

Data.md

File metadata and controls

49 lines (44 loc) · 1.63 KB
Raw

Dataset Setup for PixFoundation

PixMMVP

  • Download MMVP.
  • Download our dataset annotations from hugging face
  • Or use the referring expressions and segmentation annotations that are available under data/pixmmvp/
  • Setup the dataset directory data/PixMMVP to the following
|--- PixMMVP
   |--- MMVP Images
   |--- Questions.csv
   |--- Objects.csv
   |--- Segmentations.json
   |--- visual_patterns.csv
   |--- meta_none.txt
  • Visualizing the groundtruth annotations
python pixmmvp/dataset/test_loader.py --root data/PixMMVP/ --out_dir OUT_DIR
  • Generate the prompts variations using the script under data/

PixCV-Bench

  • Download CV-Bench
  • Use their tool to recreate the images for the 2D section only, or download directly from ADE20K and COCO2017 datasets.
  • Download our dataset annotations from hugging face
  • Or use the referring expressions and segmentation annotations that are available under data/pixcvbench/
  • Setup the dataset directory data/PixCVBench to the following
|--- PixCVBench
   |--- ADE20K
      |--- validation
         |--- images
         |--- FinalSegmentations.json
   |--- COCO
      |--- coco
         |--- images
      |--- FinalSegmentations.json
   |--- CV-Bench
      |--- test.parquet
      |--- Objects.csv
  • Visualizing the groundtruth annotations
python pixcvbench/dataset/test_loader.py --ade_root data/PixCVBench/ADE20K/ --coco_root data/PixCVBench/COCO/ --out_dir OUT_DIR