Effects of Video Compression formats on Neural Network Performance

Abstract

The field of autonomous vehicles and driverless cars is a field which makes extensive use of machine learning and artificial intelligence, relying on it to make decisions. These decisions require a vast amount of data in order to be properly inferred. This data is often in the form of images from a video feed and an increase in the amount of data is directly correlated to more refined decision making. What if you could remove certain parts of the data by compressing the image input, and how would that influence the performance of the different machine learning algorithms? In this report we have two datasets that we compress in different ways. We then analyze the results of running a pre-trained neural network model on them, and compare it’s performance to that of running the same neural network model on the non-compressed datasets. The results show that the removal of data via compression is not in a linear relationship with neural network performance, and that depending on the compression type, results may be favorable or unfavorable.