Putting AI to work in an innovative and pragmatic way
The pitfalls of AI video understanding
There are many solutions for AI video understanding yet they have crippling issues:
- They are expensive, especially when applied to large volumes of content with a low unit value.
- They output large numbers of candidate results including false-positives, which requires human selection and validation downstream.
- When false positives are not removed, the viewer experience is significantly altered (e.g. ad breaks interrupting the narrative continuity).
Despite some productivity gains, the actual field needs are not fulfilled with sufficient automation to actually enable sufficient budget savings.
Our hybrid approach reduces AI costs while augmenting the quality of results to match expectations in the field
Our approach relies on four components:
- Well-proven signal analysis methods for fundamental detections in video and audio.
- Carefully selected AI specialised AI service generating candidate data sets for editorial segmentation, diarized dialogues, etc.
- User-settable rules, related to content typology, defining the quantity and frequency of e.g. ad breaks, etc.
- Proprietary post-evaluation algorithmics: this process emulated the thought process of a skilled operator along the content flow to evaluate each and every market candidate from stages 1. & 2, so as to provide usable end results only, at the same quality as humanly produced.
Types of markers generated
Our service provides marker metadata for:
- Start and end of the useful composition.
- Start and end of titles, end credits, intro and outro.
- Ad break positions which are non narration-disruptive even when commercial blacks are not present, in exactly the required quantity as per defined rules.
Markers files at your specs
Markers are provided as an xml file or else, as per your ingest specifications.
Editorial segments are labelled as per IMF standards (SMPTE 2067-3).