Hi Adrian,
I understand that you wish to know if the average precision of your YOLOv4 detector is a relevant metric and if it’s result is good enough. Since average precision represents the accuracy of the detector and is a vector of scores for each object class in the order specified by ground truth data, 93% is a very good result. Apart from average precision there are other metrics like Recall, F1-score and Precision which can be considered. You can refer Evaluate Detection Precision for more information about calculating precision and recall.
The F1-score is the harmonic mean of precision and recall, calculated as:
F1 = 2 * (precision .* recall) ./ (precision + recall);
Apart from metrics, there are other factors which can determine if the detector has been trained well:
- Test your YOLOv4 model on a validation dataset that was not used during training. This will give you an indication of how well your model will perform on new data.
- Check if there are any specific classes that your YOLOv4 model is having trouble detecting accurately. If so, you may need to re-evaluate your annotations, gather more data, or consider using other techniques such as data augmentation or minority oversampling to improve the detection performance on those classes.
- You may also look into optimizing the hyperparameters using automatic tools like Hyperparameter Optimization in Classification Learner App.
Hope this helps!
