A client dealing with a vision system that is unable to perform predictably in adverse conditions, needs to certify that the system can work under stress. But how could they improve performance without collecting more data?
The Challenge
Discovery - which situations is the system most impacted by?
Training - what is needed to better train the mdl to handle unexpected inputs?
Deployment - what options are there to deploy the model and boost performance without significantly increasing data collection?
Advai’s Capability
Advai's library identifies the challenging scenarios that will most affect the image recognition system, and is able to to provide insight into how to improve the model.
Data augmentation and model retraining was implemented t boost performance for weak classes
Key features that distinguish the classes are deployed for critical components in the field, to boost the vision system performance.
Result
By understanding the poor performing classes, data collection was prioritised to get maximum gain for adversarial conditions.
The client was able to promote the model to production faster as they could increase the base level of performance under restrictive conditions.
As an added boost, critical assets were able to have key features amplified so that the vision system would further improve baseline performance.