<p>Zero-Shot Quantization has emerged as an efficient model compression method for deploying object detectors on resource-constrained edge devices when access to original training data is restricted. However, existing methods encounter significant challenges when applied to object detection tasks: the synthetic data fails to capture the complex spatial layouts and multi-scale object distributions in real-world detection scenarios, and the quantization training process lacks instance-aware differentiation between key objects and background noise, leading to inefficient distillation and severe performance degradation. In this paper, we propose a novel zero-shot quantization framework designed for object detection. Specifically, we first introduce a Structured Scene-aware Synthesis (S²-Syn) method to generate high-quality calibration data that closely resemble real detection scenes. In the subsequent quantization-aware training process, we propose a differential knowledge distillation strategy with Instance-Guided Feature Alignment (IGFA) to enable the quantized model to focus on learning discriminative features of key object instances, effectively alleviating the performance degradation caused by low-bit quantization. Extensive experiments across mainstream CNN-based and Transformer-based detectors demonstrate that the proposed method achieves remarkable performance. Notably, when quantizing YOLOv5-s to 6-bit on the MS COCO benchmark, the proposed method outperforms LSQ trained with the full 120k real dataset by 1.3% mAP using only 2k synthesized data, and surpasses LSQ trained on 2k real data by 3.9% mAP.</p>