stumash / northern-lights-vision

Find out if the northern lights are currently visible around the world

northern-lights-vision

Find out if the Aurora Borealis is currently visible!

• notification service (email or text or mobile app)
• live web page updated in real time

Data Pipeline

1. download_all_ylknf_480p_mp4.sh : get all the raw night sky video from ucalgary, put it in /data.northernlights.vision/unlabelled. Then we flattened the folder structure to only contain the 480p mp4 files
2. put raw night sky videos in s3, run /labeller project as s3 static site, GUI to label timeranges of raw video as 'containing northern lights'. This gave us /data.northernlights.vision/annotations json files that represent the time ranges in each raw video where the aurora is visible
3. /videoToImages contains the code of an aws lambda that (in massive parallel for $2.50) uses FFMPEG to extract all the individual frames of every video. These folders full of individual frame .jpgs are zipped at .7z compressed folders in /data.northernlights.vision/images
4. /makeLabelledVideoFramesCsv/main.py generates /data.northernlights.vision/labelledVideoFrames.csv which is a .csv file of the form File Name,Frame Time,Is Aurora which labels every zipped frame as True or False

TODO

• Assess image quality across resolutions

• • Download same-day video in three formats (480p ~50MB, 720p ~100MB, 1080p ~250MB)
• • Convert mp4 to list of png
• • Assess quality of pngs of different formats
• • Convert different png resolution to each other (up/down sample) and assess quality
• • FINDING: 480p is more than enough resolution

• Download all 2014-2018 aurora cam streams (480p)

• Flatten S3 folder structure

• Build the data-labelling pipeline

• • mp4 to png list (script)
• • label png list using timestamps for mp4 (script)
• • build ui for easy timestamping of mp4
  • • URGENT: label enough data

• Build and Train a computer vision machine learning model

• • Define the model (probably AlexNet variant)
• • Train the model (checkpoints w/ accuracies)

• Service that watches live aurora cam, feeds it to model

• • log model predictions w/ persistence service

• Build web page and notification system

• • ?