Watchlist autolearn
Watchlist Autolearn is a feature developed to considerably increase the accuracy of identification of people registered in watchlists. This feature can be mostly used in the access control use case, where people registered in watchlists are recurring periodically, usually on a daily basis.
Watchlist Autolearn every day automatically selects a face image from all the matches of a person against a watchlist member and adds the image to the corresponding watchlist member.
The face images are collected over a user-defined time period. At the end of this time period (the default value is 30 days), the oldest image added to the watchlist member is replaced by a new image from the last day. This means that over this time period, watchlist members accumulate multiple images of themselves which are updated periodically and which represent their current appearance.
The feature ensures that SmartFace Platform can match a person with the corresponding watchlist member with a higher accuracy. In addition, matching isn’t influenced by changes in the face of a person, as the collected images reflect the current face of the matched person.
Selection of the face image added to the watchlist member is based on the selection threshold. Watchlist Autolearn adds only a face with a matching score equal or higher than the selection threshold to the watchlist.
Autolearn face clustering
To increase the positive impact of the Watchlist Autolearn, faces selected daily are added into the separate collections. These collections are called clusters. Currently two clusters are supported:
- No mask cluster
- Face mask cluster
Matched faces where a face mask is present are added to the face mask cluster. Matched faces where face mask is not present are added to the no-mask cluster. Due to having the autolearn faces added into multiple clusters SmartFace Platform can optimize the selection of the faces on a per-cluster basis and also optimize the face matching.
Configuration of Watchlist Autolearn
By default, this feature is disabled and can be configured by WatchlistAutoLearnConfig configuration. The following table describes parameters of the configuration.
| Configuration property | Default Value | Description |
|---|---|---|
Enabled | false | A flag which indicates whether the Watchlist Autolearn features should be started at the defined ExecutionStartTime |
ExecutionStartTime | null | Time of day (UTC) when Watchlist Autolearn runs. The format is hour:minute:second. For example: 23:00:00. |
SelectionThreshold | 50 | The minimal threshold for the Selection strategy. A higher threshold can decrease the chance of adding an incorrect face image to the watchlist member. A lower threshold can cause watchlist poisoning, when a face image which doesn't belong to the watchlist member might be added. We recommend setting the value higher than your matching threshold. The value is used for no mask cluster. |
MaskedSelectionThreshold | 70 | The minimal threshold for the Selection strategy. A higher threshold can decrease the chance of adding an incorrect face image to the watchlist member. A lower threshold can cause watchlist poisoning, when a face image which doesn't belong to the watchlist member might be added. We recommend setting the value higher than your matching threshold. The value is used for face mask cluster. |
MaxAutoLearnFacesCount | 30 | The maximum number of Watchlist Autolearn faces that are stored for one watchlist member into one cluster. Only one autolearn face is added per day. |
NoFaceMaskConfidenceThreshold | -3000 | Required face mask confidence for the face to be selected into the No face mask cluster. The detected face needs to have a FaceMaskConfidence lower than this value. |
FaceMaskConfidenceThreshold | 3000 | Required face mask confidence for the face to be selected into the Face mask cluster. The detected face needs to have a FaceMaskConfidence higher than this value. |
Matched faces with FaceMaskConfidance value between configured NoFaceMaskConfidenceThreshold and FaceMaskConfidenceThreshold (default -3000 to 3000) are not added to the selection by the autolearn.