Overview

This dataset features 1,000 short movie clips, each paired with detailed text descriptions in a video-text modality. Sourced from the Hollywood2 dataset (as introduced in the paper below), it focuses on diverse human actions in realistic movie scenes. The 1,000 clips were curated through a structured process to ensure relevance and clarity. This collection is ideal for Generative AI (GenAI) applications like video-to-text generation, text-to-video synthesis, and multimodal content creation, while also supporting action recognition, video analysis, and related tasks in computer vision and NLP.

Dataset Details

• Video Specifications:
  • Duration: Around 20 seconds on average for efficient processing.
  • Quality: Realistic movie quality with natural settings, no artificial filters, capturing authentic human movements.
  • Camera Angle: Varied angles typical of film footage to highlight actions and contexts.
• Annotations: Text descriptions detailing actions, movements, and scene context for paired learning.
• Size: 1,000 videos selected from the original Hollywood2 collection.
• Format: Standard formats for ML pipeline integration.

Potential Applications

• Generative AI Focus:
  • Fine-tuning models for video captioning.
  • Generating narrative descriptions from visuals.
  • Text-based video retrieval.
  • Creating synthetic movie content via multimodal learning.
• Other Uses:
  • Action detection in computer vision.
  • Scene understanding and behavior analysis in video analytics.
  • Benchmarking on realistic movie data for research.

Inspired by the Hollywood2 dataset, this resource empowers GenAI innovation in movie-related AI.

---

Citation

If using this dataset, please cite: @InProceedings{marszalek09, author = "Marcin Marsza{\l}ek and Ivan Laptev and Cordelia Schmid", title = "Actions in Context", booktitle = "IEEE Conference on Computer Vision & Pattern Recognition", year = "2009" }