README.md

MoviePuzzle Dataset

Introduce

This dataset is based on MovieNet for MoviePuzzle task. We use 228 movie to generate 10031 movie clips. 7048 clips for train, 589 for val, 1178 for in domain test and 1196 for out domain test.

Download

You can download the full dataset here: https://movienet.github.io/

Structure

We categorized a dataset of 10031 movie clips into labels ranging from 0 to 10030. Then, we divided this dataset into four subsets: the training set, validation set, in-domain testing set, and out-of-domain testing set. The out-of-domain testing set consists of movies that are different from those in the other sets, while the remaining movies were split into sets with almost equal proportions. The file structure is as follows:

.
├── train/
│   ├── 1/
│   │   ├── 1_0.png
│   │   ├── 2_1.png
│   │   ├── ...
│   │   ├── subtitle.json
│   │   ├── info_suffled.json
│   │   └── info.json
│   ├── 2/
│   │   ├── 2_0.png
│   │   ├── 2_1.png
│   │   ├── ...
│   │   ├── subtitle.json
│   │   ├── info_suffled.json
│   │   └── info.json
│   └── ...
│
├── split_ clip_id.json
├── test_in_domain/
├── test_out_domain/
├── val/
└── README.md

The PNG images under the folder named clip_id/ (like 1/, 2/ ... ) represent the sampled frames in sequential order from the movie sequence. The subtitle.json file corresponds to the image frames, providing subtitles for each frame. The info.json file includes labels for each frame such as shot and scene information. On the other hand, info_shuffled.json represents the labels after shuffling. split_clip_id.json is a dictionary containing all the clip_id present in the datasets.

File format

info.json

{
    "tt_id" : "tt0047396",
    "img_num" : 20
    "img_id" : [0, 1, 2 ...],
    "shot_id" : [13, 13, 14 ...]
    "scene_id" : [6, 6, 6, ...]
}

The tt_id indicates which movie the clip belongs to. The img_num represents the number of frames that make up the image, while img_id indicates the index of the image frame. The shot_id represents the position of each image frame in its original movie sequence, and the scene_id represents the position of the image frame in its original movie scene.

In the info_shuffled.json file, the dictionaries have the same keys as in the info.json file, but their values lists have been shuffled in the same order.

subtitle.json

[
    ["Men, are you over 40?"],
    ["When you wake up in the morning, do you feel tired and rundown?"],
    ...
]

Criteria

1. The frames have subtitle in a clip is more than 80%.
2. The length of a clips is 10 to 20.
3. Only the subtitle that locate on the frame will be record.
4. NO overlap

Citation

If you find our dataset helpful, please kindly cite us in your research.