Rockchip Object Detection
YOLO-family detectors on the Rockchip NPU: conversion, post-processing and FPS.
YOLOv5, YOLOv8 and newer YOLO variants are the most common NPU workloads. The RKNN Model Zoo provides export scripts that move NPU-unfriendly post-processing out of the graph, plus C and Python demos with reference performance for each SoC.
Key points
- Export with the Model Zoo's recommended settings; the stock export often includes ops that fall back to CPU.
- Do NMS and box decoding on the CPU (it is cheap) and keep the NPU graph pure convolution.
- Batch camera frames through RGA for resize/letterbox to keep the CPU free.