← 返回模型目录

检测

YOLO11n-OBB 旋转框检测

航拍场景的旋转框检测，DOTA 15 类（船只、飞机、储罐等）。

部署到自己的板卡
QCS6490
28.5msW8A8
QCS8550
19.3msW8A16
IQ9
适配中
IQ8
19.4msW8A16

机房真机实测，全部精度 ›

运行效果

样例输入
ship 85
ship 85
ship 83
ship 83
ship 83
ship 83
ship 83
ship 83
ship 83
ship 83
ship 83
ship 82
ship 82
ship 82
ship 82
ship 82
ship 82
ship 82
ship 82
ship 82
ship 82
ship 82
ship 82
ship 82
ship 82
ship 82
ship 82
ship 80
ship 80
ship 80
ship 80
ship 80
ship 80
ship 80
ship 80
ship 80
ship 80
ship 80
ship 78
ship 78
ship 78
ship 78
ship 78
ship 78
ship 78
ship 78
ship 78
ship 78
ship 77
ship 77
ship 77
ship 77
ship 77
ship 77
ship 77
ship 77
ship 77
ship 77
ship 77
ship 77
ship 77
ship 75
ship 75
ship 75
ship 75
ship 75
ship 75
ship 75
ship 75
ship 75
ship 75
ship 75
ship 75
ship 74
ship 74
ship 74
ship 74
ship 74
ship 74
ship 74
ship 72
ship 72
ship 72
ship 72
ship 72
ship 72
ship 71
ship 71
ship 71
ship 71
ship 71
ship 71
ship 71
ship 71
ship 71
ship 69
ship 69
ship 69
ship 69
ship 69
ship 69
ship 69
ship 69
ship 69
ship 67
ship 67
ship 67
ship 67
ship 67
ship 67
ship 66
ship 66
ship 66
ship 66
ship 66
ship 66
ship 66
ship 66
ship 66
ship 64
ship 63
ship 63
ship 63
ship 63
ship 63
ship 63
ship 63
ship 63
ship 63
ship 61
ship 61
ship 60
ship 60
ship 58
ship 58
ship 58
ship 56
ship 56
ship 55
ship 53
ship 53
ship 52
ship 47
ship 45
ship 42
ship 41
ship 41
ship 41
ship 36
ship 36
ship 36
ship 35
ship 35
ship 35
ship 33
ship 33
ship 31
ship 31
ship 31
ship 28
ship 28
ship 27
ship 25
ship 25
机房真机 QCS8550 · W8A16 · 处理耗时 19.3 ms

板卡实测

真机处理时延的中位数。多阶段模型计入各阶段处理，生成模型的耗时受输入和输出长度影响。

板卡精度速度FPS峰值内存
QCS6490W8A8默认28.5 ms35.1113 MB
W8A1639.1 ms25.6116 MB
QCS8550W8A16默认19.3 ms51.9114 MB
FP1626.6 ms37.6118 MB
IQ9QCS9075W8A16适配中——
FP16适配中——
IQ8QCS8275W8A16默认19.4 ms51.5214 MB
FP1627.3 ms36.7216 MB

部署指南

  1. 1

    在工作流中跑通

    用样例或自己的输入在机房的QCS8550 上运行，确认结果和速度。

  2. 2

    导出工程

    在工作流右上角选择「导出」，下载工程包。模型与样例按 sha256 锁定，前后处理与网站运行时一致。

  3. 3

    在板卡上安装运行库

    把工程拷到QCS8550（ARM64、Python 3.10 的 AidLux 环境，AidLux ≥ 2.0），在工程目录下载并校验已验证的配套运行库。

    python3 install-sdk.py
  4. 4

    只需要模型文件？

    各板卡的构建也发布在 Hugging Face ↗，按芯片下载，文件与工程包中的完全一致。

    hf download APLUX-Agora/YOLO11n-obb --include "targets/QCS8550/*" --local-dir ./resources
  5. 5

    运行

    结果写入 outputs/，并打印每块积木的耗时。工程通过官方 AidLite 接口加载模型，代码完整随包提供。

    python3 app/pipeline.py --sdk-root .sdk --input 你的文件
AidLux 开发文档 ↗

规格

任务
检测
输入尺寸
1024×1024
来源
Model Farm
许可证
Apache-2.0
版本
1.0.0

使用它的原型

全部 1 个原型 →
浏览用到它的原型 →