Identifying and visualising the main themes emerging from a video collection of videos.

RQ: Which are the main themes (based on number of scenes) in the Amazon Fires related YouTube videos?

VISUALIZATION

Responsive image
STEPS
WHAT’S IT FOR
TOOLS
DETAILS AND MATERIALS
SCRAPING
DATA EXPLORATION
DATA PREPARATION
URLS CREATION
RENAMING THE NEW COLUMN
DOWNLOAD VIDEOS
COLLECTING VIDEOS IN A NEW FOLDER
FRAME EXTRACTIONS BY CHANGE OF SCENE
COLLECTING ALL THE FRAMES IN FOLDERS DIVIDED BY VIDEO
EXTRACTING CHARACTERS WITH RUNAWAY ML
ANNOTATE THE VISUALISATION
GET A LIST OF VIDEOS FOR EACH CHOSEN QUERY AND SELECTED TIME-FRAME.
OPEN THE YOUTUBE DATA TOOL(YDT) CSV DOWNLOADED AND EXPLORE THE DATA.
FILTER THE LIST BY ORDER OF VIEWS TO TAKE THE FIRST 10 VIDEOS AS SAMPLES.
INSIDE THE YDT.CSV THERE IS ONLY THE VIDEOS ID, BUT YOU NEED THE URL TO DOWNLOAD THEM.
TO KEEP TRACK OF THE NEW COLUMN IN WHICH WE HAVE ALL THE VIDEO URLS.
DOWNLOAD THE VIDEO SAMPLE QUICKLY AND AUTOMATICALLY.
IT’S IMPORTANT FOR THE NEXT SCRIPT THAT THE FOLDER CONTAINS ONLY THE DOWNLOADED VIDEOS.
* THE SCRIPT EXTRACTS THREE FRAMES EVERY SCENE CHANGE.
THE NEXT STEP REQUIRES HAVING THE FRAMES OF EACH VIDEO IN A SINGLE FOLDER.
INSERT THE FRAMES TO RUN THE MODEL AND INDICATE THE DIRECTORY WHERE THE NEW FRAMES WILL BE SAVED.
CREATE VIDEOS GRID FOR EACH QUERY.
Youtube Data Tools ⟶ [Video List]
Excel ⟶ [Import Data]
Excel ⟶ [Filter-Discending]
Excel ⟶ =CONCATENA(E2;F2)
Excel
Python3 ⟶ [PyTube3]
No tool needed
Python3 ⟶ [PySceneDetect]
No tool needed
DeepLabV3 ⟶ Extract people from images
Figma
“Amazon Fires” - “Pray for Amazonia”
videoIdvideoTitlepublishedAtviewCountposition

E2 ⟩ http://www.youtube.com/watch?v=

F2 ⟶ videoId

videoUrl

LINK TO PYTHON3 DOCUMENTATION

LINK TO PYTUBE3 DOCUMENTATION

LINK TO REPOSITORY AND STEP-BY-STEP GUIDE

Rename the videos inside the folder like this:

vid1vid2vid3vid4

LINK TO PYSCENEDETECT DOCUMENTATION

LINK TO REPOSITORY AND STEP-BY-STEP GUIDE

Rename the frames inside the folder like this:

frame1frame2frame3frame4frame5

LINK TO DOWNLOAD

LINK TO STEP-BY-STEP-GUIDE

LINK TO DOWNLOAD FIGMA

METODOLOGY

aim

This method aims to identify which are the main themes emerging within a collection of videos. Frame extraction for this purpose is based on scene change detection, so that the images to be analysed are only taken once and there are no duplicates due to scene length. The layout used to arrange the frames according to their visual similarity is offered by Pixplot, which uses UMAP projection, a dimensionality reduction algorithm, specifically designed for visualising complex data in low dimensions (2D or 3D).

output

The final visualisation is a clusterisation of frames sorted by visual similarity that allows the identification of predominant thematic clusters within the analysed video collection. The thematic annotations of the visualisation were drawn following the boundaries identified by the original Pixplot visualisation.