A selection for editors
Around half a million people are registered in AIDA Cruises' guest community. They upload photos of their trips. Some of this material is suitable for highlighting in the community or using for marketing.
Our image assessment considers criteria for this selection. Around one in seven images is presented to the editorial team for a decision. Aesthetic criteria play a part; submission does not automatically imply a suspected rule violation.
What the figures describe
The editorial team confirms nineteen out of twenty submitted images unchanged. This statement applies to the submitted selection. It does not describe measured quality across all other assessments.
This distinction is crucial to interpreting the results. Agreement on selected cases answers a different question from reviewing all material. Anyone using a figure must state which selection and decision it refers to.
From keywords to broader assessment
Our earlier image recognition supplied keywords from a cloud service. Today, assessment runs on our own hardware and considers more complex subjects. This is an enhanced function, not merely a change of hosting location.
We can refine the criteria and control where images are processed. This combination creates practical freedom: the technical foundation can be adapted to the task.
Operations are part of the decision
A local model needs reliable operations. An external service needs suitable integration. Cost per operation is only part of the comparison. Desired quality, data flows and the ability to change models or criteria later are just as important.
Not every task calls for our own hardware. The operating model should meet requirements and allow future changes. First, the task must be described precisely enough.
Distinguish image assessment from automated moderation
The existing image assessment and our moderation platform Modary are separate systems. Editorial confirmation of selected images is not a quality metric for Modary or evidence of savings already achieved there.
Modary is in testing. It addresses another part of the task: how a model's assessment becomes a permitted action under defined rules, and how human corrections are considered. This distinction keeps the product's status and operational experience clear.