Real-time Object Detection on Edge

YOLOv8 distilled to a 6MB student that runs at 38 FPS on Jetson Orin Nano.

0.683

mAP@0.5

26ms

Latency

6MB

Model size

The problem

Edge hardware cannot run the full teacher network in real-time, but accuracy must remain useful.

Feature-level + response-level distillation, INT8 quantization, and TensorRT graph optimization.

mAP@0.5 dropped only 2.1 points while throughput went from 9 FPS → 38 FPS on Orin Nano.