Compression of Multiscale Features of FPN with Channel-Wise Reduction for VCM

With the development of deep learning technology and abundance sensors, machine vision applications that utilize vast amounts image/video data are rapidly increasing in autonomous vehicle, video surveillance smart city fields. However, achieving a more compact representation lower latency solutions is challenging for such machine-based applications. Therefore, it essential to develop efficient coding standard Currently, Moving Picture Experts Group (MPEG) developing new called machines (VCM) with two tracks, each mainly dealing compression input (Track 2) features extracted from 1). In this paper, an enhanced multiscale feature (E-MSFC) method proposed efficiently compress generated by pyramid network (FPN), which backbone networks specified VCM evaluation framework. The E-MSFC reduces channels be included single map compresses using versatile (VVC), latest standard, rather than stream (SSFC) module existing MSFC. addition, performance further adding bottom-up structure fusion (MSFF) module, performs channel-wise reduction E-MSFC. Experimental results reveal significantly outperforms image anchor BD-rate gain up 85.94%, includes additional 0.96% achieved MSFF structure.