An experimental evaluation of image input degradation on machine learning performance
Examensarbete för masterexamen
Computer systems and networks (MPCSN), MSc
Significant advancements in the field of machine learning have been made during the past decade due to artificial neural networks being feasible to compute. To train these neural networks, a large amount of data is needed, especially in the field of autonomous driving where image data needs to be stored at a large scale. Such data can be reduced in size significantly by using lossy video compression at the cost of losing visual fidelity. This trade-off could potentially be balanced such that the data size of a dataset is reduced while the reduction in data quality is not significantly affected. This thesis aims to establish the effects of video compression on machine learning algorithms performing computer vision tasks, specifically for autonomous driving. This involved evaluating the effect of certain encoders and coding parameters on a set of ML-algorithms. Multi-objective optimisation was performed to find sets of optimum coding parameters for each encoder evaluated, referred to as coding parameter sweet spots. Experiments were conducted to measure the impacts of video compression, using the optimum coding parameters’, on machine learning algorithms. The experiment results indicated that some encoders’ sweet spots were able to compress the data without significantly altering ML-performance. Collected experiment data was also used to compare encoders capabilities and behaviour when compressing data for ML usage. Finally, suggestions for how practitioners should evaluate and validate lossy video compression methods were provided.
Machine learning , video compression , optimisation , codec , encoder , parameters , autonomous driving