Tips

For a Classic Classifier, you could say the more data you have the better.
The less bias you have in your data, the better.
Try to record data which is close to your real life use case. If your real life use case includes many different varieties, you should include these varieties in the data collection process. e.g. different light conditions, different backgrounds, different angles, when recording e.g. the same objects.
Collect separate data for test sources. We are automatically splitting out test data from the train data you are collecting if you are uploading data to sources with the Linked Test Data option enabled. However, this test data is not as separated visually from train data as it can be. The best is if you create a new test source and collect the data separately in order to have a benchmark test dataset. The closer to reality / to your use case the test data is, the better. If you have a good test dataset, you can save yourself a lot of time, since the test accuracy for the trained models on the platform will be closer to the actual performance of the SDK in real life and you can detect models that do not meet your accuracy requirements before even packaging them into an SDK.

SDK / App:

design the UX of your app in such a way that will help users in cases when AI is not working as good as expected.

Last updated 1 year ago

For a Classic Classifier, you could say the more data you have the better.
The less bias you have in your data, the better.
Try to record data which is close to your real life use case. If your real life use case includes many different varieties, you should include these varieties in the data collection process. e.g. different light conditions, different backgrounds, different angles, when recording e.g. the same objects.
Collect separate data for test sources. We are automatically splitting out test data from the train data you are collecting if you are uploading data to sources with the Linked Test Data option enabled. However, this test data is not as separated visually from train data as it can be. The best is if you create a new test source and collect the data separately in order to have a benchmark test dataset. The closer to reality / to your use case the test data is, the better. If you have a good test dataset, you can save yourself a lot of time, since the test accuracy for the trained models on the platform will be closer to the actual performance of the SDK in real life and you can detect models that do not meet your accuracy requirements before even packaging them into an SDK.

SDK / App:

design the UX of your app in such a way that will help users in cases when AI is not working as good as expected.

Last updated 1 year ago