Improved accuracy
All of the v2.0 release notes apply to v2.1, especially these:
- Model output format is NOT backwards compatible with v1
- Bind-mount required in docker-compose.yml file
Trained with additional data
Feedback from testing the v2.0 model indicated too many false positives and misclassifications. To fix this behavior, v2.1 training used more images without objects and this improved the accuracy of the model.
Recommend a LOWER threshold for Frigate object detection
For example, if Frigate running on v1 or v2 of this YOLO model used a threshold of 0.7, you may want to use a threshold value of 0.6 with v2.1 to have a similar relative rate of false positives and false negatives. See /benchmark/RESULTS.md for more information
New "animal" object class
There is a new label file that adds an object class "animal" trained using mostly small mammals: squirrels, martens, raccoons, etc.
