Developing methods to evaluate the trustworthiness of CV models

This project aims to develop new methods for evaluating reliability of CV models beyond standard metrics such as mAP, Precision, Recall, etc. In practice models with high score often shows unstable behavior in real-world situations: unstable detections, artifacts and others. As a result, a team of developers have to review a lot of data to understand if model really can be trusted before using it in real applications. Our R&D team with experience in Computer vision will study a new ways to measure models reliability and stability. The project will produce a practical evaluation methodology, research results and tools that help developers understand when AI systems can be trusted and ease human review if it needed.