Search papers, labs, and topics across Lattice.
This paper explores knowledge distillation (KD) to improve the performance of quantized object detection models for vulnerable road user (VRU) safety on edge devices. They train a compact YOLOv8-S student using a YOLOv8-L teacher, achieving 3.9x model compression while maintaining quantization robustness. Results on BDD100K show that the KD student retains accuracy under INT8 quantization, even surpassing the teacher's FP32 precision, demonstrating that KD transfers precision calibration rather than raw detection capacity.
Quantization crushes large object detection models for edge deployment, but knowledge distillation can resurrect them, even surpassing their original floating-point precision in a much smaller package.
Deploying accurate object detection for Vulnerable Road User (VRU) safety on edge hardware requires balancing model capacity against computational constraints. Large models achieve high accuracy but fail under INT8 quantization required for edge deployment, while small models sacrifice detection performance. This paper presents a knowledge distillation (KD) framework that trains a compact YOLOv8-S student (11.2M parameters) to mimic a YOLOv8-L teacher (43.7M parameters), achieving 3.9x compression while preserving quantization robustness. We evaluate on full-scale BDD100K (70K training images) with Post-Training Quantization to INT8. The teacher suffers catastrophic degradation under INT8 (-23% mAP), while the KD student retains accuracy (-5.6% mAP). Analysis reveals that KD transfers precision calibration rather than raw detection capacity: the KD student achieves 0.748 precision versus 0.653 for direct training at INT8, a 14.5% gain at equivalent recall, reducing false alarms by 44% versus the collapsed teacher. At INT8, the KD student exceeds the teacher's FP32 precision (0.748 vs. 0.718) in a model 3.9x smaller. These findings establish knowledge distillation as a requirement for deploying accurate, safety-critical VRU detection on edge hardware.
