Skip to content

M-VAD Names Dataset. Multimedia Tools and Applications (2019)

License

Notifications You must be signed in to change notification settings

aimagelab/mvad-names-dataset

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

13 Commits
 
 
 
 
 
 
 
 
 
 

Repository files navigation

M-VAD Names Dataset

This repository contains the latest version of the M-VAD Names Dataset, introduced in the following papers:

Stefano Pini, Marcella Cornia, Lorenzo Baraldi, Rita Cucchiara
Towards Video Captioning with Naming: a Novel Dataset and a Multi-Modal Approach
International Conference on Image Analysis and Processing, 2017

Stefano Pini, Marcella Cornia, Federico Bolelli, Lorenzo Baraldi, Rita Cucchiara
M-VAD Names: a Dataset for Video Captioning with Naming
Multimedia Tools and Applications, 2019

Dataset Overview

We collect and release a new set of annotations for the Montreal Video Annotation Dataset (M-VAD).

The dataset contains the annotations of characters' visual appearances, in the form of tracks of face bounding boxes, and the associations with characters' textual mentions, when available. We detect and annotate the visual appearances of characters in each video clip of each movie through a semi-automatic approach. The released dataset contains more than 24k annotated video clips, including 63k visual tracks and 34k textual mentions, all associated with their character identities.

The dataset can be downloaded here.
By downloading the dataset, you declare that you will use it for research and educational purposes only, any commercial use is prohibited.

Please cite with the following articles:

@article{pini2019mvad,
  title={{M-VAD Names: a Dataset for Video Captioning with Naming}},
  author={Pini, Stefano and Cornia, Marcella and Bolelli, Federico and Baraldi, Lorenzo and Cucchiara, Rita},
  journal={Multimedia Tools and Applications},
  volume={78},
  number={10},
  pages={14007--14027},
  year={2019}
}
@inproceedings{pini2017towards,
  title={{Towards Video Captioning with Naming: a Novel Dataset and a Multi-Modal Approach}},
  author={Pini, Stefano and Cornia, Marcella and Baraldi, Lorenzo and Cucchiara, Rita},
  booktitle={International Conference on Image Analysis and Processing},
  year={2017}
}

Please note that the original M-VAD video clips are not included in this dataset and can be downloaded from the official M-VAD website.

Technical details

The dataset file is a pickle object containing a python dictionary, structured as follows:

mvad-names.pkl
 └── <MOVIE>
      ├── 'characters'
      |    └── <CHARACTER>
      |         └── 'DVS<CLIP-ID>'  -> tracks
      └── 'videos'
           └── 'DVS<CLIP-ID>'
                └── <CHARACTER>  -> tracks

where tracks is a list of bounding boxes containing the face of the character in the form (frame-id, x_min, y_min, x_max, y_max).

Dataset splits

Original M-VAD training, validation, and test sets are obtained by splitting the 92 movies in three disjoint parts. However, when considering characters' identities, video clips of the same movie have to be in every split, so that the algorithms can learn the visual appearance of the characters on the training set and apply it on the validation and test set.

We release official training, validation, and test set for the M-VAD Names dataset which tend to respectively have 80%, 10%, and 10% of video clips of each movie, of video clips of each character, of video clips with one mention, and of video clips with two or more mentions.

Overall Avg. per movie
Train videos 19,023 207
Validation videos 2,976 32
Test videos 2,836 31

Official splits are available, along with the M-VAD Names dataset, in the Release section.

Dataset Visualization

We provide a visualization tool to show the dataset annotations.

Please install the required packages with:

pip install -r requirements.txt

Visualize the entire dataset:

python visualize.py --mvad_path <M-VAD-path> --mvad_names_path <M-VAD-Names-path>

where <M-VAD-path> corresponds to the M-VAD dataset folder which contains the video clips in a specific folder for each movie and <M-VAD-Names-path> corresponds to the M-VAD Names file path.

Visualize a specific clip of a movie:

python visualize.py --mvad_path <M-VAD-path> --movie <movie-folder> --clip <clip-ids>

Save the results in place of visualizing them:

python visualize.py --mvad_path <M-VAD-path> --save_results

For instance, the following code saves the visualizations of the clips DVS123, DVS198 of the movie IRON_MAN_2:

python visualize.py --mvad_path ./M-VAD --mvad_names_path ./mvad-names.pkl --movie IRON_MAN_2 --clip "123 198" --save_results

Refined M-VAD Captions

Along with the M-VAD Names dataset, we release an extended version of the original M-VAD movie descriptions. In particular, during the annotation process, we found that several annotated characters were not tagged as "someone" in the original M-VAD captions but were mentioned with their proper names. To fix this problem, we add new annotations (i.e. new "someone" tags) in every movie caption for each mentioned character that is not annotated in the original M-VAD, but that we have correctly annotated in the previous stage of the process. Overall, we fix 1,253 M-VAD descriptions by adding 116 unique characters that appeared in the original captions but that were not tagged as "someone".

Refined M-VAD captions can be downloaded in the Release section.