dataset.md

Dataset Preparation

• Osprey-724K 🤗 download

Data	Size
osprey_short_form.json	57 MB
osprey_conversation.json	106 MB
osprey_detail_description.json	63.4 MB
osprey_part_level.json	153 MB
osprey_lvis_positive_negative.json	140 MB

• COCO: train2017, imgs should contain all the images including training set and validation set.
• pascal_part: train.json, VOCdevkit.
• partImagenet: train_format.json, PartImageNet_OOD.
• refcocos: refcoco, refcoco+.
• vg: vg_train_with_mask.json (mask is generated from HQ-SAM), images can be downloaded from OpendataLab, image should contain all the vg images(VG_100K and VG_100K_2).
• vcr: vcr.

After downloading all of them, organize the data as follows in ./data,

├── coco
│   ├── annotations
│   │   └── instances_train2017.json
│   └── imgs
├── part data
│   ├── pascal_part
│   │   ├── train.json
│   │   └── VOCdevkit
│   └── partImagenet
│       ├── train_format.json
│       └── train
├── refcocos
│   ├── finetune_refcoco_train_with_mask.json
│   └── finetune_refcoco+_train_with_mask.json
├── Osprey-724K
│   ├── osprey_short_form.json
│   ├── osprey_conversation.json
│   ├── osprey_detail_description.json
│   ├── osprey_part_level.json
│   └── osprey_lvis_positive_negative.json
├── vg
│   ├── vg_train_with_mask.json
│   └── image
└── vcr
    ├── train.jsonl
    └── vcr1images