Advanced Applications - Optimize Model with Indicators

Optimization models are a big topic. This article describes the metrics data available on some platforms for evaluating model performance, as well as common optimization operations.

Average detection rate

After model training is completed, you can view the average detection rate of each version of the model in the "Training" page.

You can observe the performance of the model by averaging the detection rate. The average detection rate reflects the average level of recognition for each label.

Further click on the "Average detection rate", you can view the detection rate and accuracy of each label, and you can see which labels are better identified and which labels have problems with recognition. Only when the detection rate and accuracy are good, the model performs well.

For labels that are slightly worse in recognition, some hints are given (yellow exclamation mark in following picture).

Furthermore, you can click "View" to view which targets are correctly identified, which targets are missed, and which targets are incorrectly identified in the sample map, so as to guide the sample labeling, sample preparation, etc.

Attention:

• The detection rate only represents the effect of the model on the labeled image. For example, when the difference between the test image and the labeled image is large, the detection effect may not be good.
• If the average detection rate is low, the offline test will be less effective. Please ensure that the online indicators are good first, and then conduct offline testing.

Display the Confidence

Confidence, which shows which target labels were identified and with what degree of confidence, on a "Tested" sample image.

You can display the confidence on top of the image by turning on "Show Confidence" under the "More" TAB.

Confidence represent can depend on the extent of the forecast results, value range of [0, 1]. The closer the value is to 1, the more reliable the result is.

Above is an example of detecting whether the object is a "car" with a confidence of , which means that the selected object is almost certain to be a "car".0.98

When the detection target is more complex, it is not necessary to excessively pursue a high value of confidence. The confidence measure needs to be used in conjunction with a detection threshold to filter out what should be detected and what should not.

Call prediction function in the SDK, you can configure the detection threshold. See in SDK tutorial document (C++ version for example)