Real-Time-Inference-for-Unmanned-Ground-Vehicles-Using-Lossy-Compression-and-Deep-Learning

The aim of this study is to explore the effect compressed training images have on the performance of deep learning segmentation architectures and determine if lossy compression is a practical solution for providing real-time transfer speed for autonomous vehicle perception systems. As a result, this study found JPEG to achieve the highest compression ratio of 144.49× at JPEG quality level 0; while also achieving the fastest transfer speed of the compressors used on the Nvidia Xavier Edge Device. Furthermore, JPEG achieved the highest mIoU accuracy for both architectures tested in comparison to SZ3 and ZFP. Of the two deep learning architectures tested, EfficientViT outperforms U-Net for all lossy compressors at all levels of compression. EfficientViT achieves a peak mIoU of 95.5% at a JPEG quality level of 70. While U-Net peaks with an mIoU of 90.683% at a JPEG quality of 40.