MAJOR CHANGES
FASTER AND MORE ACCURATE
Typical detection speeds should be around 6ms, which is faster than the 9ms of the model from v1.x . Accuracy metrics are improved versus v1.x models.
** NOT BACKWARDS COMPATIBLE !!**
This release introduces a new model that is NOT compatible with the earlier 1.x releases. Its output is a simplified set of coordinates instead of the more complex DFL style box representation. This simplifies and accelerates post-processing on the CPU.
You must bind-mount the detector plugin code to use the new model!
Slim YOLO v9 "s" architecture
The model reduces the head width from 128 to 64 classes. This is the primary cause for the faster operation, and does not appear to degrade detection accuracy.
Additional training data
Adds data from two public data sets that have images similar to surveillance cameras. The data is cropped using a method similar to how Frigate prepares image data for object detection.