<p>Knowledge distillation plays an increasingly important role in enhancing lightweight object detection. However, traditional methods tend to neglect instance-level knowledge migration, resulting in sub-optimal performance. In this work, we propose hierarchical instance distillation (HID), a novel approach focusing on the extraction and distillation of instance knowledge. First, we obtain instance knowledge based on the teacher-student model by integrating teacher-student predictions, ensuring that the distilled knowledge retains critical instance-level information. Then, we perform hierarchical alignment of spatial scales for this knowledge through hierarchical spatial pooling and channel equalization, enabling the student model to learn multi-scale instance representations effectively. Finally, we impose a learnable mask on each instance knowledge to filter negative knowledge, avoiding the transfer of misleading or redundant information. Additionally, we utilize top-k matching for teacher-student instance knowledge, ensuring proximity-based alignment to reduce the teacher-student gap. We conduct extensive experiments with different types of detectors on the COCO dataset. Our method achieves significant over 3 AP improvements and consistently outperforms other knowledge distillation methods, demonstrating its effectiveness and robustness in enhancing lightweight object detection.</p>