Why?
METATAGGER.IO is the solution.
Revolutionizing the process of AI-driven subject identification and imagetagging METATAGGER.IO delivers an unprecedented level of accuracy, speed and value maximizing through an unsupervised, self-learning, face clustering algorithm developed on top of AWS Rekognition.
What?
METATAGGER.IO was developed for Woodwing Elvis as a SaaS and can easily be expanded to other DAM / CMS systems. The integration points are simple and isolated parts of the system.
The fully automated integration with DAM (Elvis) via S3 and REST delivers a typical time from an image is added to Elvis to metadata is updated with identified profiles of below 20 seconds.
METATAGGER.IO requires no modification of existing DAM or cloud or on-premise solution. Using a single-point system agnostic API enabling integration into virtually any system METATAGGER.IO extracts for analysis and returns IDs enriched with all desired and required metatags needed for identification, classification, organising, use and reuse.
METATAGGER.IO is built on top of AWS Rekognition as an unsupervised, self-learning, face clustering algorithm using a serverless version of DBScan.
How
METATAGGER.IO tackles some of the key challenges in managing and processing digital assets.
Duplicates
When the same physical person is assigned multiple clusters. In order to handle this METATAGGER.IO has features allowing easy reassignment to either an existing person or an entirely new person. Subsequently METATAGGER.IO ensures that similar duplicates are automatically corrected.
False positives
Face detection has false positives. We have seen faces detected on pizzas and other obvious false positives. The METATAGGER.IO face detection algorithm has been expanded to allow thresholds to be specified in order to reduce the noise.
Domino effect of false positives
When managing and procession an ever-expanding digital stock, persons may be assigned to wrong clusters. Often very unsharp or non-camera-facing faces get grouped together. Given enough images this cluster grows to be best match for faces that ideally should have been assigned to a different existing profile.
METATAGGER.IO solves this by enabling setting thresholds for both quality of facial capture, threshold for how large a part of the face is shown in the image etc. to reduce the false positive domino effect of low-quality facial recognitions.
Balancing false positives and duplicate person clusters
A generic problem to all face-clustering algorithms is how the distance sensitivity settings affect the balance between false positives and duplicate person clusters. The custom algorithms used in METATAGGER.IO provide several different methods to mitigate this problem. Furthermore, the system provides several advanced methods for editors to adjust and assist the AI in these matters.
Access and usability
Central to METATAGGER.IO is our high level of UX. Developed specifically to allow non-technical users to work with naming of identified persons and handling of false positives and duplicates. The UI allows for advanced users to tweak every aspect of the face-clustering algorithms and how the output is pushed back to the DAM, and still providing simple, image based workflows for all editorial staff working with naming identified persons. In this process both advanced users and editorial staff participate in optimizing and training the system.