Skip to content

Mean average precision and more: how to train and evaluate a vision model

For Computer vision teams and students10 stagesOutline: full hub coming
The short answer

Training a vision model starts with a clean train and test split and ends with honest evaluation: precision and recall, the F1 score, a confusion matrix and mean average precision (mAP) for detection. Compare versions on the same held-out images before you switch.

01Split train and test

02Augment the data

03Object detection vs image classification

04Train

05Precision, recall and F1

06Read a confusion matrix

07Mean average precision (mAP)

08Compare versions

09Test on new images

10Decide what to fix next