While it doesn't provide a huge mAP50-95, its predictions are always correct on the files I've tested. Confidence could be higher, but the model is very precise.
Confusion matrix of the model, generated on the test set
PR curve of the model, generated on the test set
Other metrics are available in the ./demo/val subfolder.
Demo
Manual annotation
yolov8_animeface
Comparison with an existing model
While we can argue about the comparison between two models from different generations, I believe it is interesting to compare this model with zymk9's model, based on yolov5. They both have the same purpose and were trained in the same fashion.
On the same dataset with the same parameters (conf=0.001 & iou=0.6), yolov8x6-animeface produces better metrics than the one on yolov5x.
Images Instances Box(P R mAP50 mAP50-95):
1003 1566 0.778 0.685 0.633 0.232
Speed: 1.5ms pre-process, 85.5ms inference, 1.4ms NMS per image at shape (32, 3, 1280, 1280)
yolov5-anime provides better results when images are resized at 640px, but it still is inferior to yolov8-animeface with the same parameters.
Surprisingly enough, yolov5 is way more confident that yolov8. However, it also has way more false positives.
Yolov8_animeface
Yolov5_anime
Conclusion
While this model performs greatly on my dataset, my metrics can differ from yours. One lighter model might be released on a later occasion.
