How we use Serverless & Step Functions to quickly adapt our workflows disguised with a buzzwordy title with examples of Machine Learning and cloud architecture
Presented at MetEXPO (SMPTE) Sydney 18th July 2019
43. “
Original ResizeEnergy map Seams
Energy map edge detection
Algorithm finds important parts
Generates individual seams (top to down, or left to right)
Compute the cumulative minimum energy connected seams
Content aware image resizing
48. “
7plus built a face detection tool:
● Face recognition algorithm
● Train with dozens of cast images (from IMDB)
● Find safest crop point
DIY Machine Learning Face Detection Tool
50. “
● Used by Google, Yahoo, Microsoft, Intel, IBM, Sony, Honda, Toyota
● Trained and deployed in many use cases:
● Stitching street view images
● Detecting surveillance video
● Helping robots navigate
● Inspecting labels on products
● Rapid face detection ← Cool!
Open Source Computer Vision Library
61. News & Talk Shows
Clearly enunciated
No background sound
Regular presenters
Premium Drama
Various countries
Range of accents
Different actors
62. News & Talk Shows
Clearly enunciated
No background sound
Regular presenters
Premium Drama
Various countries
Range of accents
Different actors
Fishing Shows
Presented outside
Strong accents
Slang & species
68. “Windsor” - Tow (0.4940) wins (0.9490) a (0.9801) bridge (0.6395)
“Quay” - Circular (1.0000) key (0.3497)
“How the fight” - About (0.9996) hell (0.8830) the (0.9879) f--k (0.8509)
“Turnbull” - Alleged (0.9986) malcolm (0.9986) terrible (0.4550) also (1.0000)
Each translation is scored, these
indications used for review
69. Why 7plus reviews output?
1. As a broadcaster we’ve committed to highest
accuracy on our content
2. When publishing news, we could be liable
when an error implicates others
70. Developers praise the technology;
but considered user reaction
"Are they illiterate?, this is embarrassing!"
72. “
Humans are uniquely adaptable
Humans have so much other contextual knowledge not available to a machine transcription algorithm
Humans are much better transcribers in difficult situations, poor quality audio, speakers are far from recording devices, or
speakers are familiar with one another.
Machine translation has similar human-level performance. So essentially it is a propagation of error problem.
Doing machine translation on a machine transcription will have the same performance as if you used human translation on the
text from a human transcription
78. “
Group rows into sentence
1. The company already employs
2. many Australians
3. in its four other
4. overseas offices.
VTT language is segmented into a
broken format, we group logically first
Results
La compañía ya emplea a muchos
australianos en sus otras cuatro
oficinas en el extranjero.
Das Unternehmen beschäftigt
bereits viele Australier in seinen
vier anderen Auslandsbüros.