metadata
dataset_info:
features:
- name: image_id
dtype: string
- name: image
dtype: image
- name: prompt_open
dtype: string
- name: prompt_close
dtype: string
- name: objects
dtype: string
- name: relationships
dtype: string
splits:
- name: train
num_bytes: 807872243
num_examples: 5000
download_size: 781301448
dataset_size: 807872243
task_categories:
- image-text-to-text
license: apache-2.0
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
This dataset, derived from VG150, provides image-text pairs for scene graph generation. Each example includes an image, an "open" prompt, a "close" prompt, a list of objects, and their relationships. It's designed to be used for training and evaluating models that generate scene graphs from images and textual prompts.
This dataset is used in the paper R1-SGG: Compile Scene Graphs with Reinforcement Learning. Please replace PLACEHOLDER_PAPER_LINK with the actual link once available.
The dataset is structured as follows:
- image_id: Unique identifier for the image.
- image: The image itself.
- prompt_open: An open-ended prompt related to the image.
- prompt_close: A more specific prompt related to the image.
- objects: A list of objects present in the image.
- relationships: A description of the relationships between the objects.
Data Usage:
The dataset can be loaded using the datasets library:
from datasets import load_dataset
db_train = load_dataset("JosephZ/vg150_train_sgg_prompt")["train"]
db_val = load_dataset("JosephZ/vg150_val_sgg_prompt")["train"]
(Further instructions from the original README regarding training and inference can be included here)