I’m loving this feature. Makes me wonder how it’s implemented. Did a Google search, found nothing.
Obviously it’s doing some image analysis based on the video frame. Is it though? Is Amazon using info from IMDB? The latter’s also owned by Amazon.
I think these kind of algorithms are fucking cool.
Any thoughts/ideas?
#entertainment #media
Want to see the real deal?
More inside scoop? View in App
More inside scoop? View in App
blind
SUPPORT
FOLLOW US
DOWNLOAD THE APP:
FOLLOWING
Industries
Job Groups
- Software Engineering
- Product Management
- Information Technology
- Data Science & Analytics
- Management Consulting
- Hardware Engineering
- Design
- Sales
- Security
- Investment Banking & Sell Side
- Marketing
- Private Equity & Buy Side
- Corporate Finance
- Supply Chain
- Business Development
- Human Resources
- Operations
- Legal
- Admin
- Customer Service
- Communications
Return to Office
Work From Home
COVID-19
Layoffs
Investments & Money
Work Visa
Housing
Referrals
Job Openings
Startups
Office Life
Mental Health
HR Issues
Blockchain & Crypto
Fitness & Nutrition
Travel
Health Care & Insurance
Tax
Hobbies & Entertainment
Working Parents
Food & Dining
IPO
Side Jobs
Show more
SUPPORT
FOLLOW US
DOWNLOAD THE APP:
comments
First component is face recognition based on image analysis. There are two opportunities and data sources for this analysis. One is the source video that's decoded during transcoding, and the other are the JPEG thumbnails extracted during transcoding for purposes of seeking/navigation. Scene change detection can also be performed during transcoding which makes it possible to segment the timeline into scenes.
The other key component is IMDB. It provides you with both a cast list for the movie as well as photos of most of the cast.
So once you have a cast list and their photos, performing facial recognition isn't very difficult because you're working with a constrained set of possible faces to match. The facial recognition algorithm doesn't need to be perfect, it just needs to be good enough to pick the likeliest match out of a constrained set of cast photos.
An alternative solution (or just a method of increasing facial recognition likelihood) would be to run speech-to-text conversion of the audio track (the one featuring the original cast) and then match it up to a script or caption file that associates dialog with character names. That would also cover cases where an actor is only heard but not seen (e.g. voiceover) which otherwise wouldn't work with only facial recognition.
But it's also entirely possible they just pay people somewhere to watch the movie and index all the characters/actors manually. For companies which already do subtitling/captioning this extra task wouldn't be much of a stretch. 🙂