Fregata 0.17.2.5

Built on Frigate v0.17.2

v0.18.0-beta3 is also available here: https://github.com/3rdBitLabs/Fregata/releases/tag/v0.18.0-beta3

What's Changed

🚀 CoreML Engine Re-Write for Frigate AI Enrichments Performance

In this release we've completely re-written Fregata's CoreML Engine to better support Frigate's AI Enrichments such as Semantic Search, License Plate Recognition, and Facial Recognition. All GPU-eligible models now run 100% on your Mac's GPU bringing huge performance increases when running these models.

Performance Gains vs v0.17.2.4 on a Base M4 Mac Mini

Model	Performance Gain
Semantic Search
Jinav1 (Large) - Text	~419%
Jinav1 (Large) - Vision	~273%
Jinav2 (Large) - Text	~442%
Jinav2 (Large) - Vision	~341%
License Plate Recognition
YOLOv9 Plate Detection	~231%
PaddleOCR Text Detection (Small, single-line plates)	~161%
PaddleOCR Text Detection (Large, multi-line plates)	~181%
PaddleOCR Text Recognition	~188%
Facial Recognition
ArcFace (Large)	~411%

*Performance increase will vary depending on your Mac model. Pro, Max, and Ultra chips should do especially well here due to having more and/or better GPU cores.

Due to this change, we've altered the default model size that is loaded vs Frigate Docker. If you have Semantic Search or Facial Recognition set to model_size: small we recommend removing that line in your config, or setting it to model_size: large for both better accuracy and better performance. For Semantic Search and Facial Recognition, the small models can ONLY run on the CPU and are less accurate, for LPR both small and large run on the GPU but large is only useful for multi-line license plates and is less accurate on single-line plates.

New Defaults:

semantic_search:
  model_size: large

face_recognition:
  model_size: large

lpr:
  model_size: small #LPR large should only be used for countries with multi-line license plates

---

⚙️ Improvements & Changes

• The minimum macOS version is now 14. This was necessary to support new CoreML APIs for the AI Enrichment performance improvements that do not exist on macOS 13.

• Make the CPU stats in the bottom left of the Frigate WebUI show Fregata's CPU usage, instead of the Mac's total system CPU usage.

• Make the RAM usage in the Fregata menu bar more accurate. It was undercounting RAM for some users, and overcounting for others.

• Improve crash telemetry so we can fix more reported crashes.

• Timeout moving clips from the RAM disk to the Media Directory after 30s, so a stalled copy does not halt all clips copying.

🐛 Bug Fixes

• Fix an incorrect log statement if the RAM disk gets too full.

• Fix a bug that could cause clip preview files to not be saved if the Media Directory has a space in its path.

• Fix a crash seen in telemetry that could cause Frigate to crash if a camera name was longer than 27 characters.

• Fix a crash seen in telemetry that could cause Frigate to crash in rare cases when ONVIF auto tracking was enabled.

• Fix a crash seen in telemetry that could cause process to crash in rare cases during Fregata shutdown, instead of exiting gracefully.

• Fix a crash seen in telemetry in sqlite-vec that could happen in rare cases while deleting embeddings from the vector database by bumping the sqlite-vec version.

• Fix a crash seen in telemetry that could occur in rare cases when RecordingCleanup is running.

• Fix a crash seen in telemetry around shared memory access that was very rare.

Download

• Fregata 0.17.2.5 for macOS (Apple Silicon)

SHA-256: 2991df897d726725c7de01590605fc979fe187ffdd7bbc4607c1a763d98346c6

Requires macOS 14 (Sonoma) or later on Apple Silicon.