Video Localized Narratives

Visit the project page for all the information about Video Localized Narratives, data downloads, data formats, visualizations, and more. The arXiv version of our CVPR paper can be found here.

Here we provide the official code for the publication "Connecting Vision and Language with Video Localized Narratives". The code can be used to load and visualize the Video Localized Narrative (VidLN) annotations. Additionally, we provide code to evaluate the tasks of Video Narrative Grounding (VNG) and Video Question-Answering (VideoQA) with the sub-tasks of text-output questions and location-output questions.

Abstract

We propose Video Localized Narratives, a new form of multimodal video annotations connecting vision and language. In the original Localized Narratives, annotators speak and move their mouse simultaneously on an image, thus grounding each word with a mouse trace segment. However, this is challenging on a video. Our new protocol empowers annotators to tell the story of a video with Localized Narratives, capturing even complex events involving multiple actors interacting with each other and with several passive objects. We annotated 50k videos of the OVIS, UVO, Oops, and Kinetics datasets, totalling 3.5M words. Based on this data, we also construct new benchmarks for video narrative grounding and video question-answering tasks, and provide reference results from strong baseline models. Our annotations are available at https://google.github.io/video-localized-narratives/.

Setup

Installation

See install.md for how to install the Python dependencies.

If when trying to run the code you get an error like

ModuleNotFoundError: No module named 'video_localized_narratives'

it most liekly means that you didn't add the video-localized-narratives folder to your PYTHONPATH.

Data Preparation

See data_preparation.md for instructions on how to prepare the data.

Video Localized Narratives (VidLNs)

Interactive Python Demo

From the video-localized-narratives folder run

python3 video_localized_narratives/tools/demo.py

to see an interactive visualization of sample VidLN annotations.

HTML Visualizer

To visualize Video Localized Narratives interactively in the web-browser, open video-localized-narratives/local_html_vidln_viewer/index.html with your web-browser.

You will have to select the jsonl file to load and the data root from which the frames, audio files, and video files will be loaded.

For the data root select video-localized-narratives/data/. The web browser will show a message like "This will upload all files from "data". Only do this if you trust this site." Click on "Upload" here.

Note that this will not actually upload any data, as the whole website runs locally on your own PC. However, this is necessary to give the local website access to the data folder. The data folder has to have the correct folder, i.e. the same as the sample data supplied in this repository with folders frames, recordings, and videos (where videos and recordings are optional) and in each of these folders a sub-folder with the name of the dataset (e.g. OVIS_train) is expected.

Afterwards, select the VidLN jsonl file, for example video-localized-narratives/data/vidlns/OVIS_train_sample.json Then you should see a visualization of a sample VidLN annotation, you can play the audio and see how the mouse cursor moves interactively. You can also hover with the mouse over a word to see the mouse trace for this word.

Video Narrative Grounding (VNG)

Interactive Demo

From the video-localized-narratives folder run

python3 video_localized_narratives/video_narrative_grounding/demo.py

to see an interactive visualization of sample VidLN annotations. Note that for this to work, you first have to download some data (see above).

Evaluation

You can run the evaluation for VNG like this

python3 video_localized_narratives/video_narrative_grounding/eval_vng.py --meta_filename=data/vng/OVIS_VNG/meta_expressions/test/meta_expressions.json --extra_masks_filename=data/vng/OVIS_VNG/extra_masks/test/extra_masks.json --orig_masks_filename=data/vng/OVIS_VNG/orig_masks/annotations_train.json --result_folder=/path/to/your/vng_result/

The orig_masks_filename has to point to the json annotation file of the original dataset. We do not provide this, you have to download it from the websites of the original datasets (OVIS and UVO). Note that we sub-split the original OVIS training set into a training and a test set for VNG. For both VNG sub-splits, the orig_masks_filename has to point to the original annotations for the training set.

The folder with results, e.g. /path/to/your/vng_result/ has to contain sub-folder for each video, with sub-folders for each expression id, that contain png files with masks for each frame.

For example, the result for OVIS should look like this

/path/to/your/ovis_vng_result/
├── 028f6f64
│   ├── 0
│   │   │── img_0000001.png
│   │   │── ...
│   │   │── img_0000036.png
│   ├── 1
│   │   │── img_0000001.png
│   │   │── ...
│   │   │── img_0000036.png
│   ├── 2
│   │   │── img_0000001.png
│   │   │── ...
│   │   │── img_0000036.png
│   ├── 3
│   │   │── img_0000001.png
│   │   │── ...
│   │   │── img_0000036.png
├── 11e16068
│   ├── 0
│   │   │── img_0000001.png
│   │   │── ...
│   │   │── img_0000067.png
│   │── ...
...

And your result for UVO should look like this

/path/to/your/uvo_vng_results/
├── 00jZej9_xh8
│   ├── 0
│   │   │── 0.png
│   │   │── ...
│   │   │── 300.png
│   ├── 1
│   │   │── 0.png
│   │   │── ...
│   │   │── 300.png
│   ├── 2
│   │   │── 0.png
│   │   │── ...
│   │   │── 300.png
├── 00MUo_0F9-Q
│   ├── 0
│   │   │── 0.png
│   │   │── ...
│   │   │── 301.png
│   │── ...
...

Video Question-Answering (VideoQA)

Here we explain how to evaluate video question-answering task on the Oops-QA benchmark. The benchmark is split into the text-output and location-output tasks. The total score for the Oops-QA benchmark is obtained by averaging the results for the two sub-tasks.

Text-output

Evaluate the Oopa-QA text-output task using

python3 video_localized_narratives/videoqa/text_output/eval_text_output.py --gt_json_path=data/videoqa/text_output/oops_val/qa_text_output.json --results_path=/path/to/your/results.json

Where results.json should have the following format:

{
  "question_val_0": "black shorts", 
  "question_val_1": "trampoline",
  ...,
  "question_val_12416": "riding",
}

Location-output

Evaluate the Oopa-QA location-output task using

python3 video_localized_narratives/videoqa/location_output/eval_location_output.py --gt_json_path=data/videoqa/location_output/oops_val/qa_location_output.json --result_folder=/path/to/your/results/

In the result folder, e.g. /path/to/your/results/, you should have one folder for each video, which in turn should have one sub-folder for each question_hash, which then contains png files with the predicted masks for every frame (the structure is very similar to the VNG task).

Your result for the Oops-QA validation set should look like this

/path/to/your/results/
├── 34 Funny Kid Nominees - FailArmy Hall Of Fame (May 2017)10
│   ├── 69e69b3a8b29d92dd94be66c79f82d69
│   │   │── 000000.png
│   │   │── ...
│   │   │── 000218.png
├── 34 Funny Kid Nominees - FailArmy Hall Of Fame (May 2017)11
│   ├── 92dcc24e4c843c13fc5f8902a4e10a8a
│   │   │── 000000.png
│   │   │── ...
│   │   │── 000316.png
│   ├── a424642d0f7769403c66c69de1944757
│   │   │── ...
│   ├── ...
...

Citation

If you use this code for a publication, please cite

@inproceedings{Voigtlaender23CVPR,
  author        = {Paul Voigtlaender and Soravit Changpinyo and Jordi Pont-Tuset and Radu Soricut and Vittorio Ferrari},
  title         = {{Connecting Vision and Language with Video Localized Narratives}},
  booktitle     = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year          = {2023}
}