Pixel-level attention based data compression for spike camera

Abstract With the rapid development of autonomous driving applications, there is an increasing demand for high-speed vision sensors, like spike cameras. However, the data transmission and storage requirements are becoming increasingly burdensome due to the high temporal resolution, such as 40,000Hz. To resolve this problem, we propose a pixel-level attention-based data compression method for the spike camera. Firstly, the input spike data is partitioned into two block types by pixel-level attention-based method. Then, the two blocks are condensed using different methods, and side information marking is transmitted for spike decoding. Finally, the condensed spike and side information marking are compressed into a binary stream for storage. Experimental results show that our method achieves higher compression efficiency than conventional methods. The decompressed spike can also reconstruct the image with better visual quality.