alitourani/Popcorn_Dataset

🍿 Popcorn Dataset

This dataset contains visual features obtained from a wide range of movies (full-length), their extracted shots, and free trailers. It contains frame-level extracted visual features and aggregated version of them. Popcorn can be used in recommendation, information retrieval, classification, etc tasks.

📃 Table of Content

• How to Use
• Dataset Stats
• Files Structure

🚀 How to Use?

Dataset Web-Page

Check the detailed information about the dataset in its web-page presented in the link in https://recsys-lab.github.io/popcorn_dataset/.

The Designed Framework for Benchmarking

In order to use, exploit, and generate this dataset, a framework titled Popcorn is implemented. You can read more about it on the GitHub repository.

📊 Dataset Stats

General

Aspect	Value
Total number of movies	274
Average frames extracted per movie	7,732
Total number of frames/embeddings	2,158,301
Total number of full-movie frames/embeddings	2,118,647
Total number of trailer frames/embeddings	39,654

Hybrid (combined with MovieLens 25M (link) with sampling 25%)

Aspect	Value
Average movie ratings:	3.88/5
Total users (|U|):	32,663
Total items (|I|):	255
Total interactions (|R|):	413,493
|R| / |U|:	12.66
|R| / |I|:	1621.54
Sparsity:	95.04%

Required Capacity

Data	Model	Total Files	Size on Disk
Full Movies	incp3	84,872	35.8 GB
Full Movies	vgg19	84,872	46.1 GB
Movie Shots	incp3	16,713	7.01 GB
Movie Shots	vgg19	24,598	13.3 GB
Trailers	incp3	1,725	681 MB
Trailers	vgg19	1,725	885 MB
Aggregated Full Movies	incp3	84,872	10 MB
Aggregated Full Movies	vgg19	84,872	19 MB
Aggregated Movie Shots	incp3	16,713	10 MB
Aggregated Movie Shots	vgg19	24,598	19 MB
Aggregated Trailers	incp3	1,725	10 MB
Aggregated Trailers	vgg19	1,725	19 MB
Total	-	214,505	~103.9 GB

🗄️ Files Structure

Level I. Primary Categories

The dataset contains six main folders and a stats.json file. The stats.json file contains the meta-data for the sources. Folders 'full_movies', 'movie_shots', and 'movie_trailers' keep the atomic visual features extracted from various sources, including full_movies for frame-level visual features extracted from full-length movie videos, movie_shots for the shot-level (i.e., important frames) visual features extracted from full-length movie videos, and movie_trailers for frame-level visual features extracted from movie trailers videos. Folders 'full_movies_agg', 'movie_shots_agg', and 'movie_trailers_agg' keep the aggregated (non-atomic) versions of the described items.

Level II. Visual Feature Extractors

Inside each of the mentioned folders, there are two folders titled incp3 and vgg19, referring to the feature extractor used to generate the visual features, which are Inception-v3 (GoogleNet) and VGG-19, respectively.

Level III. Contents (Movies & Trailers)

A: Atomic Features (folders full_movies, movie_shots, and movie_trailers)

Inside each feature extractor folder (e.g., full_movies/incp3 or movie_trailers/vgg19) you can find a set of folders with unique title (e.g., 0000000778) indicating the ID of the movie in MovieLens 25M (link) dataset. Accordingly, you have access to the visual features extracted from the movie 0000000778, using Inception-v3 and VGG-19 extractors, in full-length frame, full-length shot, and trailer levels.

B: Aggregated Features (folders full_movies_agg, movie_shots_agg, and movie_trailers_agg)

Inside each feature extractor folder (e.g., full_movies_agg/incp3 or movie_trailers_agg/vgg19) you can find a set of json files with unique title (e.g., 0000000778.json) indicating the ID of the movie in MovieLens 25M (link) dataset. Accordingly, you have access to the aggregated visual features extracted from the movie 0000000778 (and available on the atomic features folders), using Inception-v3 and VGG-19 extractors, in full-length frame, full-length shot, and trailer levels.

Level IV. Packets (Atomic Feature Folders Only)

To better organize visual features, each movie folder (e.g., 0000000778) has a set of packets named as packet0001.json to packet000N.json saved as json files. Each packet contains a set of objects with frameId and features attributes, keeping the equivalent frame-ID and visual feature, respectively. In general, every 25 object (frameId-features pair) form a packet, except the last packet that can have less objects.

The described structure is presented below in brief:

> [full_movies]    ## visual features of frame-level full-length movie videos
  > [incp3]        ## visual features extracted using Inception-v3
    > [movie-1]
      > [packet-1]
      > [packet-2]
      ...
      > [packet-m]
    > [movie-2]
    ...
    > [movie-n]
  > [vgg19]        ## visual features extracted using VGG-19
    > [movie-1]
    ...
    > [movie-n]
> [movie_shots]    ## visual features of shot-level full-length movie videos
  > [incp3]
    > ...
  > [vgg19]
    > ...
> [movie_trailers] ## visual features of frame-level movie trailer videos
  > [incp3]
    > ...
  > [vgg19]
    > ...
> [full_movies_agg] ## aggregated visual features of frame-level full-length movie videos
  > [incp3]         ## aggregated visual features extracted using Inception-v3
    > [movie-1-json]
    > [movie-2]
    ...
    > [movie-n]
  > [vgg19]         ## aggregated visual features extracted using VGG-19
    > [movie-1]
    ...
    > [movie-n]
> [movie_shots_agg] ## aggregated visual features of shot-level full-length movie videos
> [movie_trailers_agg]    ## aggregated visual features of frame-level movie trailer videos

stats.json File

The stats.json file placed in the root contains valuable information about the characteristics of each of the movies, fetched from MovieLens 25M (link).

[
  {
        "id": "0000000006",
        "title": "Heat",
        "year": 1995,
        "genres": [
            "Action",
            "Crime",
            "Thriller"
        ]
    },
    ...
]

📚 Citation

@article{tourani2025rag,
  title={RAG-VisualRec: An Open Resource for Vision-and Text-Enhanced Retrieval-Augmented Generation in Recommendation},
  author={Tourani, Ali and Nazary, Fatemeh and Deldjoo, Yashar},
  journal={arXiv preprint arXiv:2506.20817},
  year={2025}
  doi={https://doi.org/10.48550/arXiv.2506.20817}
}

@article{villammbench,
  title={ViLLA-MMBench: A Unified Benchmark Suite for LLM-Augmented Multimodal Movie Recommendation},
  author={TBD},
  journal={TBD},
  year={2025}
}