// the find
jakowenko/double-take
Unified UI and API for processing and training images for facial recognition.
Double Take is a single Docker image that puts one UI, REST API and MQTT interface in front of several face recognition backends: CompreFace, Amazon Rekognition, DeepStack and Facebox. It is aimed at Frigate and Home Assistant users who want to know which known person just passed a camera, and it works mostly as glue between those systems and the detector.
The detector abstraction is the real value. Every backend returns the same match and unknown result shape, so switching from CompreFace to Rekognition is a config change rather than a rewrite of your automations. The MQTT output is documented with actual payload examples, including per-camera topics and a person count that resets after 30 seconds, which saves guesswork. Per-camera overrides for thresholds, masks, schedules and Frigate zones mean one global config doesn't have to compromise between a doorbell and a driveway. The Frigate loop retries against latest.jpg and snapshot.jpg and stops on the first match, which is a sensible way to get past a bad frame without hammering the detector.
The last push was 2025-08-18 and the README says nothing about release cadence, so the project looks stale as of now. Check the release tags and open issues before depending on it for a security setup. Two of the four detectors are effectively orphaned: Facebox came from Machine Box, which shut down years ago, and DeepStack has been superseded by CodeProject.AI. Those config entries will not get fixes. Authentication is off by default, there is no HTTPS in the bundled setup, and anonymous telemetry is on by default. The README discloses that last one, but most people won't read that far. This is a wrapper, so accuracy, latency and false positives are whatever the chosen backend gives you. The README has no benchmarks and little guidance on tuning beyond the default thresholds.