MINOR CHANGES
Compatible with other YOLO models
Some YOLO models published elsewhere expect input to be in (signed) INT8 format, instead of the (unsigned) UINT8 that Frigate uses. This detector can now provide these models with the input data they expect. This is an unsupported capability. If you want to run a different YOLO model, this might allow you to do so at your own risk and through your own effort.
Simplified Post-Processing (no change in speed)
The optional 3rd tensor introduced in v1.5 has been deprecated. This is the result of additional performance measurement that indicated there was no reduction in detection time from the model change. The improved speed in v1.5 came from code changes, not model changes. The models that include the optional 3rd tensor will continue to be supported, but the redundant information in the 3rd tensor will be ignored.
One-shot inference
In the /benchmarks folder there is a new script to run object detection on a single image. This is useful for comparing scores among different models for a single test image.