A model with 95% accuracy on the validation dataset can fail spectacularly in the field: at night, in the rain, with dust on the lens, or camera vibration.
At ZKA we train with augmentation specific to real conditions, validate in the field before going to production, and monitor models in continuous operation. When the environment changes, we retrain with the new data captured by the system itself.
Accuracy is measured on the dataset. Robustness is measured in months of operation without intervention.