Determining Performance of Computer Vision Models using Generalizability, Robustness & Elasticity Score

Performance measurement of computer vision models provides information about their ability to classify objects. However, their performance gets affected in the real-world environment. We propose a modification for the metric called Generalizability, Robustness, and Elasticity score (GRE), which is used to determine the efficiency of the computer vision models. Specifically, we use unaltered Visual Question Answering (VQA) datasets and develop three new datasets for each attribute of the GRE score. The new datasets pass through three novel serial processes designed to enhance the quality of the datasets. The new datasets have a better distribution of feature information of the objects in the original dataset. Their performance is measured by running the datasets on three models specifically modified for our experiment. Two out of three models perform better on our new datasets and provide a better GRE score. We prove that our system works and can provide better results than the conventional method of measuring the performance of computer vision models.