A Fully 4-bit Quantized MobileNet-SSD
Aeri Kim, Yu-Mi Kim, Eunchong Lee, Minkyu Lee, Sang-Seol Lee · 2024
Quantizing the detector is more challenging than the classifier; consequently, previous studies used some layers as floating-point layers. This leads to the problem that quantized detectors cannot be accelerated in integer-only operating devices. In this study, we present a fully 4-bit quantized MobileNet-SSD. Furthermore, we propose an optimized weight freezing order and a Knowledge Distillation (KD) loss function to minimize accuracy degradation. Our detector achieves mAP 65.09% in PASCAL VOC, which is 0.63 points higher than prior work, and we conducted detailed experiments about the proposed KD loss function.