Autoencoder-based Knowledge Distillation For Quantized YOLO Detector

Object detectors like YOLO often suffer performance degradation when quantized to lower bit precision. To mitigate this, research explores knowledge distillation (KD) methods alongside quantization. This paper proposes a novel KD method using an autoencoder for quantizing YOLO. This method involves training an autoencoder on full-precision network features, effectively transferring them to the quantized network. Applied to YOLOv7-tiny with the PASCAL VOC dataset, our method improved the mAP(0.5:0.95) of LSQ and LLTQ methods by 0.6 and 0.9, respectively. Compared to the conventional KD method, we observed an improvement in mAP(0.5:0.95) of up to 0.5.