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

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