jhelsby/ovdsgg-action-genome-split

OvDSGG Action Genome Open-Vocabulary Split

This dataset repository contains the open-vocabulary category split used by OvDSGG for Action Genome experiments.

It does not redistribute Action Genome videos, frames, or full annotations. Users must obtain and process Action Genome separately, then use this split metadata to reproduce the OvDSGG open-vocabulary training/evaluation protocol.

Paper: https://huggingface.co/papers/2608.14835
Code: https://github.com/jhelsby/OvDSGG
Model checkpoints: https://huggingface.co/jhelsby/OvDSGG

Getting Action Genome

This repository provides only the OvDSGG open-vocabulary split metadata. To train or evaluate OvDSGG, download Action Genome separately, process it with the Action Genome Toolkit, and place the processed COCO-style annotation files from this folder in the annotations directory.

The expected dataset layout is:

action_genome/
  annotations/
    ag_train_coco_style.json
    ag_test_coco_style.json
    ...
  frames/
  videos/

Usage

from datasets import load_dataset

dataset = load_dataset("jhelsby/ovdsgg-action-genome-split")
split = dataset["train"]

Each row has:

• category_type: object or relation
• name: Action Genome category name
• index: zero-based index in the full OvDSGG Action Genome class list
• ov_split: base or novel

Example:

base_objects = [
    row["name"]
    for row in dataset["train"]
    if row["category_type"] == "object" and row["ov_split"] == "base"
]

Split Summary

• Objects: 25 base, 11 novel, 36 total
• Relations: 18 base, 8 novel, 26 total

The split is generated by tools/propose_ag_split.py in the OvDSGG repository and is hardcoded in datasets/ag.py for model training and evaluation.

Citation

@misc{ovdsgg2026,
  title={OvDSGG},
  year={2026},
  url={https://huggingface.co/papers/2608.14835}
}