Dataset for Evaluating the Effectiveness of AI-Generated Data for Video-Based Action Recognition

=======================
License
=======================
This dataset is made available under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.

=======================
Summary
=======================
DATASET FOR EVALUATING THE EFFECTIVENESS OF AI-GENERATED DATA FOR VIDEO-BASED ACTION RECOGNITION

This repository contains the dataset accompanying the paper "Evaluating the Effectiveness of AI-Generated Data for Video-Based Action Recognition". The dataset was curated to investigate the impact of incorporating AI-generated synthetic video data into the training process of deep learning models (such as I3D and TimeSformer) for human action recognition. It is particularly focused on evaluating cross-domain generalization and addressing the synthetic-to-real distribution shift using domain adaptation techniques.

It is obligatory to cite the following paper in every work that uses the dataset:
K. Gomulka, P. Wozniak, T. Krzeszowski: Evaluating the Effectiveness of AI-Generated Data for Video-Based Action Recognition, Electronics, MDPI, 2026.

=======================
Data description
=======================
The dataset focuses on 15 target action categories from the HMDB51 benchmark: Draw Sword, Fall, Flick Flack, Handstand, Jump, Kick, Pick, Punch, Run, Shoot Bow, Shoot Gun, Sit, Throw, Walk, and Wave.

The repository consists of three main data subsets:
* Grok Synthetic Dataset (Grok): Contains 1,500 synthetic video sequences (100 videos per class) generated using the Grok Imagine 1.0 model. Resolution is 560x560 pixels with a duration of 6 seconds per clip.
* Meta Synthetic Dataset (Meta): Contains 1,480 synthetic video sequences (80 in the 'throw' class, 100 videos per other classes) generated using the Meta AI Vibes model. Resolution is 624x624 pixels with a duration of 5 seconds per clip.
* HMDB51 Subset (hmdb51_org): A specific subset of the real-world HMDB51 dataset containing 2,315 video sequences distributed across the 15 target action classes, providing a direct real-world baseline.

The repository also includes sample .csv files defining the train, validation, and test splits. In the example configuration, 60% of the HMDB51 dataset is allocated to training, 20% to validation, and 20% to testing, with all synthetic Grok videos appended exclusively to the training set.

=======================
Dataset structure
=======================
* Grok/ - directory containing 1,500 synthetic videos generated by Grok Imagine 1.0.
* Meta/ - directory containing 1,480 synthetic videos generated by Meta AI Vibes.
* hmdb51_org/ - directory containing 2,315 real-world videos from the HMDB51 dataset.
* ghtrain_hval/ - directory containing example .csv files defining the train, validation, and test splits for the baseline HMDB + Grok setup.
    * train.csv - training set paths and labels.
    * val.csv - validation set paths and labels.
    * test.csv - testing set paths and labels.

* README.md - detailed documentation including dataset overview, configuration details, and instructions for training the TimeSformer model using the provided data splits.

=======================
Generation Methodology
=======================
The synthetic videos were generated using a structured, combinatorial prompt engineering framework resulting in 100 unique prompts per class. Each prompt combined predefined descriptions of the action, environmental context (e.g., vintage film style, urban outdoor), and camera framing to ensure high scene diversity. While the generated sequences preserve essential motion semantics, they also may contain inherent generative artifacts.