Object Detector Friendly Frame Memory Compression Based on 1-D DWT and SPIHT

Frame memory compression (FMC) based on embedded compression (EC) modules is widely used to reduce the memory bandwidth (BW) in object detectors based on convolutional neural networks (CNNs), but the accuracy drop of the object detectors is inevitable due to due to image distortion caused by lossy EC. This paper proposes an integrated platform that equips FMC based on 1-D discrete wavelet transform (DWT) and set partitioning in hierarchical trees (SPIHT) to YOLOv2, a representative object detector. Furthermore, we propose a technique to improve the compression efficiency of 1-D DWT-SPIHT in order to maximize latency and power savings by reducing memory BW while minimizing the accuracy degradation caused by FMC. In particular, while most of the existing FMC studies have focused on improving the image quality, this study has a distinction in that it has focused on the accuracy of object detectors. The proposed method reduces the memory access of the input image by 50%, improving the throughput and reducing power consumption in proportion to the compression ratio with only a 0.01% reduction in mAP compared to baseline YOLOv2 without FMC.