RAEng Industrial Fellowship: Intelligent Data Compression

SPRING

The idea

Automotive sensors produce a large volume of data. Whether it needs to be transmitted to a processing unit or to be stored, the amount of data is unsustainable. Prof Donzella was awarded a Royal Academy of Engineering Fellowship in "Smart data compression for automotive environmental perception sensors" to tackle this problem together a strong industrial collaboration with onsemi.

Object detection performance on lossy data

Lossy data compression provides much higher compression ratio at the cost of introducing minor changes to the data which cannot be recovered. The automotive KITTI dataset was compressed using different levels of AVC and HEVC compression codecs. Machine learning based object detectors (Faster-RCNN and Yolov5) were retrained with the compressed data which demonstrated that some compression can improve the detection performance, and be used as a form of fine-tuning. Additionally, the lossy compression did not adversely affect detection performance, up to a compression ratio of 160:1.

RGB vs Bayer

RGB images contain 3 colour intensity channels, as opposed to Bayer which is 1 intensity channel. Hence, RGB carries 3 times more data than Bayer. Most existing machine learning algorithms use RGB images as the input. This work explored the effect of changing the input to Bayer. A Yolov4 network was modified from the first layer to take 1 channel inputs. The results demonstrated that there were no major detriment to using Bayer data, even with a network which was not bespoke designed for Bayer inputs.