Evaluation of Interpolation Methods for Image Downsampling in Automotive Computer Vision

Achieving real-time performance is an important goal for automated driving and ADAS applications. One optimisation for such systems is the use of lower resolution images for CNN based object detection, which can greatly improve inference speed. Reducing image resolution reduces the size of the image but also reduces image quality. The downsampling method used in ADAS is a topic often not considered when downsizing images, and this study aims to address this gap. This study investigates how downsampling using different interpolation methods impacts machine vision performance. Several common machine vision algorithms are trained on downsampled data, and their performance is evaluated. The downsampling methods used are: Bilinear interpolation, Bicubic interpolation, Nearest Neighbour interpolation, Area-Based resampling and Lanczos4 interpolation. The results show that training with different downsampling methods does have a consistent impact on performance across different object detection algorithms; however, the differences are generally very small, with a difference of less than 2% AP50 in most cases. One object detection model (RT-DETR) is shown to be much more sensitive to interpolation methods. This study indicates which methods of downsampling are best suited for use in ADAS applications, and their relative advantages and disadvantages of each method. The results presented here are relevant to designers of ADAS who are concerned with real-time optimisations.