Introduction

YouTube is today the largest video sharing platform.
In 2019, it counts more than a billion of users, more than 7 billion videos and more than 300 hours of video are uploaded each minute. Nowadays, smartphones, drones, board cameras, dash camera have become common and more content is produced everyday an uploaded on the platform, making it an almost inexhaustible source of videos of our world.
But how could this impressive database be exploited?

Using photogrammetry techniques, it is possible to extract from a series of pictures, the 3D structure of the photographed scene. To do so, a set of different frames of the same subject photographed on different angles is processed and a geometric model of the subject is generated.

Numerizing the world has become one of the challenges of the 21st century, and combining YouTube and photogrammetry could be an affordable solution. Indeed, instead of having to invest huge amounts of money to numerize the world that surrounds us, why don’t we recycle the content of YouTube to generate 3D models of the world?
This questions motivated this study.

The goal was to see if the material available for free on the platform could be used to generate sparse 3D model. A positive answer implies that an enormous database representing many locations around the globe is accessible without the need of new numerization campaigns and all
the costs related. The generated models could be shared freely and could track the evolution of the surface of the globe as many places are recorded several times.

Research summary

To investigate the question of wether or not YouTube can be used a source for the generation of 3D models of the world, a draining pipeline was implemented.

Pipeline summary

The pipeline was required to takes as input a link to a YouTube video and to output the 3D models generated from the video.

The pipeline realizes the following tasks:

1. Provided a link to a YouTube video such as this footage of Lausanne shot by a drone, the video is downloaded. (realized using youtube-dl)
2. Frames are extracted at a constant rate (say 2 frames are extracted for 1 second of video). (realized using ffmpeg)
3. The extracted frames are used to reconstruct the 3D structure of the recorded scenes. (realized using the openMVG library)

When the pipelien finishes, set_X/ (X is an index) folders were generated, each one containing the frames that were used to generate a 3D sparse point-cloud model of the scene captured in the different frames.

Illustrations

Consider the following frame of Lausanne's cathedral.

Using the draining pipeline one can generate a 3D sparse model that looks like:

Where the green points are the estimated camera poses.

Installation and Usage

Dependencies:

Python

The pipeline is basically a python wrapping that orchestrates calls to the external libraries. The code is written in python 3 and installing jupyter notebook is also required to execute the notebooks provided.
Both can simply be installed by downloading the Anaconda distribution that can be found at this [link] (https://www.anaconda.com/distribution/) (select python 3.XX distribution).

External librairies

The draining pipeline uses three main components: youtube-dl, ffmpeg and openMVG.

To install the latest versions of youtube-dl and ffmpeg please refer to the following links:

• install youtube-dl
• install ffmpeg

The actual version of the pipeline uses a slightly modified version of openMVG: some lines of the source code were modified for convenience and the the pipeline uses the develop branch of openMVG (time of writting: 3rd of July 2019).

The used version is provided in the folder openmvg_folder. To install openMVG, simply follow the installations instructions provided on the openMVG github.

We recommend installing MeshLab for the visualization of the 3D models: available here

Test the installation

After installing all the dependencies, the drain_one.py script can be run to download and process the aforementioned drone footage of Lausanne. To do so, open a terminal in the scripts/ folder and execute:
$ python drain_one.py

The output of youtube-dlshould be displayed in the terminal, followed by those of the ffmpeg and the openMVG.

The executions can take several minutes to complete.
Once it finished, in the videos/ folder one can find the test_video/ folder. In test_video/ there is a folder named using the YouTube ID of the downloaded video which is the LnyKeqdzQao/ in this case.

In LnyKeqdzQao/:

• data/ contains the video file and descriptive file.
• frames/ contains the images that the program discarded and did not use in any of the generated 3D models.
• iter0/ contains files generated during the process.
• sets/ can contain one or several set_X folders. In each set_X:
  • The frames/ folder containing the frames used for the 3D reconstruction.
  • The features/ folder and the sfm_data.json file are data used by openMVG in the reconstruction.
  • The 3D model can be found in the folder incremental/ and is generically named sfm_data_color.ply for each set (can be opened using MeshLab).

If the installation and the process succeeded you should have the same point cloud model of Lausanne's cathedral as shown above.

To process another video, simply modify the variable URL in drain_one.py and re-execute the script.

Have fun !

License

MIT License

Copyright (c) 2019 Xabier Rubiato

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.