YOLO26实现安卓ncnn实战部署：支持检测、分割、姿态估计、旋转框检测

一、前言

在伦敦YOLO Vision 2025大会上，Ultralytics创始人Glenn Jocher演示YOLO26在树莓派上实时检测的效果，我才彻底被打脸。那个只有512MB内存的小设备，竟能以28FPS的速度准确识别出视频中快速移动的小鸟，而功耗仅仅0.8W！

今年1月14日，Ultralytics正式发布了YOLO26，基于官方发布的测试数据与实际验证，YOLO26在边缘计算场景的表现确实实现了显著突破。YOLO26的设计理念发生了根本转变：其设计重心从追求基准测试性能，转向优化真实场景下的部署效率与稳定性。它不再单纯追求在标准数据集上的那一两个百分点提升，而是把重心放在了如何让AI模型在真实世界中更好地工作。

二、环境搭建（Windows版）

#1. 创建conda环境
conda create -n yolo26 python=3.11.9 -y
#2. 激活环境
conda activate yolo26
#3. 安装torch和torchvision
pip install torch==2.2.2 torchvision==0.17.2 --index-url https://download.pytorch.org/whl/cu121
#4.安装ultralytics
pip install  -U ultralytics -i https://pypi.tuna.tsinghua.edu.cn/simple
#4. 预训练权重下载
wget https://github.com/ultralytics/assets/releases/download/v8.4.0/yolo26n.pt
wget https://github.com/ultralytics/assets/releases/download/v8.4.0/yolo26n-seg.pt
wget https://github.com/ultralytics/assets/releases/download/v8.4.0/yolo26n-pose.pt
wget https://github.com/ultralytics/assets/releases/download/v8.4.0/yolo26n-obb.pt
#5. 量化依赖库
pip install -U pnnx ncnn

使用指令或者创建infer.py进行推理

yolo predict model=yolo26n.pt source="https://ultralytics.com/images/bus.jpg"

from ultralytics import YOLO

# Load a model
model = YOLO("yolo26n.pt")  # load an official model

# Predict with the model
results = model("https://ultralytics.com/images/bus.jpg")  # predict on an image

# Access the results
for result in results:
  xywh = result.boxes.xywh  # center-x, center-y, width, height
  xywhn = result.boxes.xywhn  # normalized
  xyxy = result.boxes.xyxy  # top-left-x, top-left-y, bottom-right-x, bottom-right-y
  xyxyn = result.boxes.xyxyn  # normalized
  names = [result.names[cls.item()] for cls in result.boxes.cls.int()]  # class name of each box
  confs = result.boxes.conf  # confidence score of each box

三、pnnx转检测模型

1. 导出yolo26{n/s/m}.torchscript

#看个人需求，选择模型
yolo export model=yolo26n.pt format=torchscript
yolo export model=yolo26s.pt format=torchscript
yolo export model=yolo26m.pt format=torchscript

一定要将生成的.torchscript后缀的模型单独放置文件夹，后面步骤会生成比较多的文件

2. 转换静态尺寸输入的 torchscript

#看个人需求，选择模型
pnnx yolo26n.torchscript
pnnx yolo26s.torchscript
pnnx yolo26m.torchscript

在进行转换时，会出现一些警告，BinaryOp floor_divide not supported yet，BinaryOp remainder not supported yet，目前暂不支持这两个操作，经过排查，这两个操作基本上都是在后处理中，对于后面结果暂时还没有影响...

之后会生成以下多个文件

3. 修改 yolo26{n/s/m}_pnnx.py 模型脚本

以yolo26n为例，不同量级模型可能在宽度和深度上有微小区别。
1. 第一处注意力模块（高层特征）

# 原始（固定 20×20 → 400）：
# v_96 = v_95.reshape(1, 2, 128, 400)

# 改造后（动态 H×W）：
B, C, H, W = v_95.shape
v_96 = v_95.reshape(B, 2, 128, -1)

# 后续恢复空间维度时也使用 H, W：
# v_106 = v_105.reshape(1, 128, 20, 20)
# v_107 = v_99.reshape(1, 128, 20, 20)
v_106 = v_105.reshape(B, 128, H, W)
v_107 = v_99.reshape(B, 128, H, W)

2. 第二处注意力模块（更深/更大感受野）

# 原始（同样假设 20×20）：
# v_211 = v_210.reshape(1, 2, 128, 400)

# 改造后：
B, C, H, W = v_210.shape
v_211 = v_210.reshape(B, 2, 128, -1)

# v_221 = v_220.reshape(1, 128, 20, 20)
# v_222 = v_214.reshape(1, 128, 20, 20)
v_221 = v_220.reshape(B, 128, H, W)
v_222 = v_214.reshape(B, 128, H, W)

3. 检测头输出（三个尺度）------ 动态展平 + 维度调整

# Box 分支（原硬编码 6400/1600/400 = 80²/40²/20²）
v_240 = v_239.reshape(1, 4, -1).transpose(1, 2)   # 替代 .reshape(1, 4, 6400)
v_246 = v_245.reshape(1, 4, -1).transpose(1, 2)   # 替代 .reshape(1, 4, 1600)
v_252 = v_251.reshape(1, 4, -1).transpose(1, 2)   # 替代 .reshape(1, 4, 400)

# Cls 分支（同理）
v_263 = v_262.reshape(1, 80, -1).transpose(1, 2)  # 替代 .reshape(1, 80, 6400)
v_273 = v_272.reshape(1, 80, -1).transpose(1, 2)  # 替代 .reshape(1, 80, 1600)
v_283 = v_282.reshape(1, 80, -1).transpose(1, 2)  # 替代 .reshape(1, 80, 400)

# 拼接方式修正（按 anchor 数量拼接，非通道）
v_253 = torch.cat((v_240, v_246, v_252), dim=1)   # 所有 box 预测
v_284 = torch.cat((v_263, v_273, v_283), dim=1)   # 所有 cls 预测

4. 裁剪后处理，提前返回原始 logits

# 删除从 v_285 开始的所有 anchor 解码、sigmoid、topk、NMS 等逻辑
# 直接输出检测头原始结果：
v_det = torch.cat((v_253, v_284), dim=2)  # [1, N_total, 4 + num_classes]
return v_det

4. 重新导出 yolo26{n/s/m} torchscript

python -c 'import yolo26n_pnnx; yolo26n_pnnx.export_torchscript()'
python -c 'import yolo26s_pnnx; yolo26s_pnnx.export_torchscript()'
python -c 'import yolo26m_pnnx; yolo26m_pnnx.export_torchscript()'

5. 动态尺寸转换新 torchscript

pnnx yolo26n_pnnx.py.pt inputshape=[1,3,640,640] inputshape2=[1,3,320,320]
pnnx yolo26s_pnnx.py.pt inputshape=[1,3,640,640] inputshape2=[1,3,320,320]
pnnx yolo26m_pnnx.py.pt inputshape=[1,3,640,640] inputshape2=[1,3,320,320]

四、pnnx转分割模型

1. 动态注意力计算（两处）

# 第一处（v_95 → v_107）
B, C, H, W = v_95.shape
v_96 = v_95.reshape(B, 2, 128, -1)
v_106 = v_105.reshape(B, 128, H, W)
v_107 = v_99.reshape(B, 128, H, W)

# 第二处（v_210 → v_222）
B, C, H, W = v_210.shape
v_211 = v_210.reshape(B, 2, 128, -1)
v_221 = v_220.reshape(B, 128, H, W)
v_222 = v_214.reshape(B, 128, H, W)

2. 检测与 mask 系数输出重构

v_240 = v_239.reshape(1, 4, -1).transpose(1, 2)      # box
v_263 = v_262.reshape(1, 80, -1).transpose(1, 2)     # cls
v_290 = v_289.reshape(1, 32, -1).transpose(1, 2)     # mask coeff

v_253 = torch.cat((v_240, v_246, v_252), dim=1)   # [1, N_total, 4]
v_284 = torch.cat((v_263, v_273, v_283), dim=1)   # [1, N_total, 80]
v_303 = torch.cat((v_290, v_296, v_302), dim=1)   # [1, N_total, 32]
#显式构建检测头输出
v_det = torch.cat((v_253, v_284), dim=2)  # [1, N, 84]

3. Mask Prototype 上采样动态化 + 提前返回

# 动态上采样到 P3 特征图尺寸（v_170 的 H/W）
v_341 = F.upsample_nearest(v_340, size=(v_170.shape[2], v_170.shape[3]))
v_345 = F.upsample_nearest(v_344, size=(v_342.shape[2], v_342.shape[3]))

4. 关键代码修改

  # --- 分类分支 ---
  v_284 = torch.cat((v_263, v_273, v_283), dim=1) 
  # --- 构建 detection head output: [box, cls] ---
  v_det = torch.cat((v_253, v_284), dim=2) 

  v_285 = self.model_23_one2one_cv4_0_0_conv(v_170)
  v_286 = self.model_23_one2one_cv4_2_1_act(v_285)
  v_287 = self.model_23_one2one_cv4_0_1_conv(v_286)
  v_288 = self.pnnx_unique_86(v_287)
  v_289 = self.model_23_one2one_cv4_0_2(v_288)
  # v_290 = v_289.reshape(1, 32, 6400)
  v_290 = v_289.reshape(1, 32, -1).transpose(1,2)

  v_291 = self.model_23_one2one_cv4_1_0_conv(v_197)
  v_292 = self.pnnx_unique_87(v_291)
  v_293 = self.model_23_one2one_cv4_1_1_conv(v_292)
  v_294 = self.pnnx_unique_88(v_293)
  v_295 = self.model_23_one2one_cv4_1_2(v_294)
  # v_296 = v_295.reshape(1, 32, 1600)
  v_296 = v_295.reshape(1, 32, -1).transpose(1,2)

  v_297 = self.model_23_one2one_cv4_2_0_conv(v_233)
  v_298 = self.pnnx_unique_89(v_297)
  v_299 = self.model_23_one2one_cv4_2_1_conv(v_298)
  v_300 = self.pnnx_unique_90(v_299)
  v_301 = self.model_23_one2one_cv4_2_2(v_300)
  # v_302 = v_301.reshape(1, 32, 400)
  v_302 = v_301.reshape(1, 32, -1).transpose(1,2)

  # v_303 = torch.cat((v_290, v_296, v_302), dim=2)
  # --- mask coefficient 分支 ---
  v_303 = torch.cat((v_290, v_296, v_302), dim=1) 

  # --- 构建 mask prototype ---
  v_339 = self.model_23_proto_feat_refine_0_conv(v_197)
  v_340 = self.pnnx_unique_91(v_339)
  # v_341 = F.upsample_nearest(v_340, size=(80,80))
  v_341 = F.upsample_nearest(v_340, size=(v_170.shape[2], v_170.shape[3]))
  v_342 = (v_170 + v_341)
  v_343 = self.model_23_proto_feat_refine_1_conv(v_233)
  v_344 = self.pnnx_unique_92(v_343)
  # v_345 = F.upsample_nearest(v_344, size=(80,80))
  v_345 = F.upsample_nearest(v_344, size=(v_342.shape[2], v_342.shape[3]))
  v_346 = (v_342 + v_345)
  v_347 = self.model_23_proto_feat_fuse_conv(v_346)
  v_348 = self.pnnx_unique_93(v_347)
  v_349 = self.model_23_proto_cv1_conv(v_348)
  v_350 = self.pnnx_unique_94(v_349)
  v_351 = self.model_23_proto_upsample(v_350)
  v_352 = self.model_23_proto_cv2_conv(v_351)
  v_353 = self.pnnx_unique_95(v_352)
  v_354 = self.model_23_proto_cv3_conv(v_353)
  v_355 = self.pnnx_unique_96(v_354)

  return v_det, v_303, v_355

五、pnnx转姿态估计模型

1. 动态注意力模块（两处）

# 第一处（v_95）
B, C, H, W = v_95.shape
v_96 = v_95.reshape(B, 2, 128, -1)
v_106 = v_105.reshape(B, 128, H, W)
v_107 = v_99.reshape(B, 128, H, W)

# 第二处（v_210）
B, C, H, W = v_210.shape
v_211 = v_210.reshape(B, 2, 128, -1)
v_221 = v_220.reshape(B, 128, H, W)
v_222 = v_214.reshape(B, 128, H, W)

2. 检测与关键点输出结构化

# Box 分支 (4D)
v_240 = v_239.reshape(1, 4, -1).transpose(1, 2)   # [1, N, 4]
# Class 分支 (1D，姿态任务通常单类)
v_263 = v_262.reshape(1, 1, -1).transpose(1, 2)   # [1, N, 1]
# Keypoints 分支 (51D = 17 keypoints × (x, y, conf))
v_298 = v_297.reshape(1, 51, -1).transpose(1, 2)  # [1, N, 51]
v_253 = torch.cat((v_240, v_246, v_252), dim=1)  # [1, N_total, 4]
v_284 = torch.cat((v_263, v_273, v_283), dim=1)  # [1, N_total, 1]
v_303 = torch.cat((v_298, v_300, v_302), dim=1)  # [1, N_total, 51]
v_det = torch.cat((v_253, v_284), dim=2)  # [1, N, 5]
return v_det, v_303

3. 关键代码修改

v_251 = self.model_23_one2one_cv2_2_2(v_250)
# v_252 = v_251.reshape(1, 4, 400)
v_252 = v_251.reshape(1, 4, -1).transpose(1,2)
# --- box 分支---
v_253 = torch.cat((v_240, v_246, v_252), dim=1)

v_254 = self.model_23_one2one_cv3_0_0_0_conv(v_170)
v_255 = self.model_23_one2one_cv3_2_1_1_act(v_254)
v_256 = self.model_23_one2one_cv3_0_0_1_conv(v_255)
v_257 = self.pnnx_unique_75(v_256)
v_258 = self.model_23_one2one_cv3_0_1_0_conv(v_257)
v_259 = self.pnnx_unique_76(v_258)
v_260 = self.model_23_one2one_cv3_0_1_1_conv(v_259)
v_261 = self.pnnx_unique_77(v_260)
v_262 = self.model_23_one2one_cv3_0_2(v_261)
# v_263 = v_262.reshape(1, 1, 6400)
v_263 = v_262.reshape(1, 1, -1).transpose(1,2)

v_264 = self.model_23_one2one_cv3_1_0_0_conv(v_197)
v_265 = self.pnnx_unique_78(v_264)
v_266 = self.model_23_one2one_cv3_1_0_1_conv(v_265)
v_267 = self.pnnx_unique_79(v_266)
v_268 = self.model_23_one2one_cv3_1_1_0_conv(v_267)
v_269 = self.pnnx_unique_80(v_268)
v_270 = self.model_23_one2one_cv3_1_1_1_conv(v_269)
v_271 = self.pnnx_unique_81(v_270)
v_272 = self.model_23_one2one_cv3_1_2(v_271)
# v_273 = v_272.reshape(1, 1, 1600)
v_273 = v_272.reshape(1, 1, -1).transpose(1,2)

v_274 = self.model_23_one2one_cv3_2_0_0_conv(v_233)
v_275 = self.pnnx_unique_82(v_274)
v_276 = self.model_23_one2one_cv3_2_0_1_conv(v_275)
v_277 = self.pnnx_unique_83(v_276)
v_278 = self.model_23_one2one_cv3_2_1_0_conv(v_277)
v_279 = self.pnnx_unique_84(v_278)
v_280 = self.model_23_one2one_cv3_2_1_1_conv(v_279)
v_281 = self.pnnx_unique_85(v_280)
v_282 = self.model_23_one2one_cv3_2_2(v_281)
# v_283 = v_282.reshape(1, 1, 400)
v_283 = v_282.reshape(1, 1, -1).transpose(1,2)
# --- 分类分支 ---
v_284 = torch.cat((v_263, v_273, v_283), dim=1)

# --- 构建 detection head output: [box, cls] ---
v_det = torch.cat((v_253, v_284), dim=2) 

v_285 = self.model_23_one2one_cv4_0_0_conv(v_170)
v_286 = self.model_23_one2one_cv4_2_1_act(v_285)
v_287 = self.model_23_one2one_cv4_0_1_conv(v_286)
v_288 = self.pnnx_unique_86(v_287)
v_289 = self.model_23_one2one_cv4_1_0_conv(v_197)
v_290 = self.pnnx_unique_87(v_289)
v_291 = self.model_23_one2one_cv4_1_1_conv(v_290)
v_292 = self.pnnx_unique_88(v_291)
v_293 = self.model_23_one2one_cv4_2_0_conv(v_233)
v_294 = self.pnnx_unique_89(v_293)
v_295 = self.model_23_one2one_cv4_2_1_conv(v_294)
v_296 = self.pnnx_unique_90(v_295)
v_297 = self.model_23_one2one_cv4_kpts_0(v_288)
# v_298 = v_297.reshape(1, 51, 6400)
v_298 = v_297.reshape(1, 51, -1).transpose(1,2)

v_299 = self.model_23_one2one_cv4_kpts_1(v_292)
# v_300 = v_299.reshape(1, 51, 1600)
v_300 = v_299.reshape(1, 51, -1).transpose(1,2)

v_301 = self.model_23_one2one_cv4_kpts_2(v_296)
# v_302 = v_301.reshape(1, 51, 400)
v_302 = v_301.reshape(1, 51, -1).transpose(1,2)

# v_303 = torch.cat((v_298, v_300, v_302), dim=2)
#  --- 关键点 ---
v_303 = torch.cat((v_298, v_300, v_302), dim=1)

return v_det, v_303

六、pnnx转旋转框obb模型

1、 动态注意力模块（两处）

# 第一处（v_95）
B, C, H, W = v_95.shape
v_96 = v_95.reshape(B, 2, 128, -1)
v_106 = v_105.reshape(B, 128, H, W)
v_107 = v_99.reshape(B, 128, H, W)

# 第二处（v_210）
B, C, H, W = v_210.shape
v_211 = v_210.reshape(B, 2, 128, -1)
v_221 = v_220.reshape(B, 128, H, W)
v_222 = v_214.reshape(B, 128, H, W)

2、 输出结构化重构：Box + Class + Angle 解耦

# Box 分支 (4D: cx, cy, w, h)
v_240 = v_239.reshape(1, 4, -1).transpose(1, 2)   # [1, N, 4]

# Class 分支 (15D: 多类别置信度)
v_263 = v_262.reshape(1, 15, -1).transpose(1, 2)  # [1, N, 15]

# Angle 分支 (1D: 旋转角度余弦/正弦或直接角度)
v_290 = v_289.reshape(1, 1, -1).transpose(1, 2)   # [1, N, 1]

v_253 = torch.cat((v_240, v_246, v_252), dim=1)  # [1, N_total, 4]
v_284 = torch.cat((v_263, v_273, v_283), dim=1)  # [1, N_total, 15]
v_303 = torch.cat((v_290, v_296, v_302), dim=1)  # [1, N_total, 1]

v_det = torch.cat((v_253, v_284), dim=2)  # [1, N, 19] = [box(4) + cls(15)]

return v_det, v_303
# - v_det:  [1, N, 19] → (cx, cy, w, h, cls_0, ..., cls_14)
# - v_303:  [1, N, 1]  → 旋转角度（或其编码，如 cos/sin）

3、关键代码改动

v_247 = self.model_23_one2one_cv2_2_0_conv(v_233)
v_248 = self.pnnx_unique_73(v_247)
v_249 = self.model_23_one2one_cv2_2_1_conv(v_248)
v_250 = self.pnnx_unique_74(v_249)
v_251 = self.model_23_one2one_cv2_2_2(v_250)
# v_252 = v_251.reshape(1, 4, 1024)
v_252 = v_251.reshape(1, 4, -1).transpose(1,2)
# --- 合并三个尺度的 box 预测---
v_253 = torch.cat((v_240, v_246, v_252), dim=1)

v_254 = self.model_23_one2one_cv3_0_0_0_conv(v_170)
v_255 = self.model_23_one2one_cv3_2_1_1_act(v_254)
v_256 = self.model_23_one2one_cv3_0_0_1_conv(v_255)
v_257 = self.pnnx_unique_75(v_256)
v_258 = self.model_23_one2one_cv3_0_1_0_conv(v_257)
v_259 = self.pnnx_unique_76(v_258)
v_260 = self.model_23_one2one_cv3_0_1_1_conv(v_259)
v_261 = self.pnnx_unique_77(v_260)
v_262 = self.model_23_one2one_cv3_0_2(v_261)

# v_263 = v_262.reshape(1, 15, 16384)
v_263 = v_262.reshape(1, 15, -1).transpose(1,2)

v_264 = self.model_23_one2one_cv3_1_0_0_conv(v_197)
v_265 = self.pnnx_unique_78(v_264)
v_266 = self.model_23_one2one_cv3_1_0_1_conv(v_265)
v_267 = self.pnnx_unique_79(v_266)
v_268 = self.model_23_one2one_cv3_1_1_0_conv(v_267)
v_269 = self.pnnx_unique_80(v_268)
v_270 = self.model_23_one2one_cv3_1_1_1_conv(v_269)
v_271 = self.pnnx_unique_81(v_270)
v_272 = self.model_23_one2one_cv3_1_2(v_271)
# v_273 = v_272.reshape(1, 15, 4096)
v_273 = v_272.reshape(1, 15, -1).transpose(1,2)

v_274 = self.model_23_one2one_cv3_2_0_0_conv(v_233)
v_275 = self.pnnx_unique_82(v_274)
v_276 = self.model_23_one2one_cv3_2_0_1_conv(v_275)
v_277 = self.pnnx_unique_83(v_276)
v_278 = self.model_23_one2one_cv3_2_1_0_conv(v_277)
v_279 = self.pnnx_unique_84(v_278)
v_280 = self.model_23_one2one_cv3_2_1_1_conv(v_279)
v_281 = self.pnnx_unique_85(v_280)
v_282 = self.model_23_one2one_cv3_2_2(v_281)
# v_283 = v_282.reshape(1, 15, 1024)
v_283 = v_282.reshape(1, 15, -1).transpose(1,2)
# --- 合并三个尺度的 cls 预测 ---
v_284 = torch.cat((v_263, v_273, v_283), dim=1)

v_285 = self.model_23_one2one_cv4_0_0_conv(v_170)
v_286 = self.model_23_one2one_cv4_2_1_act(v_285)
v_287 = self.model_23_one2one_cv4_0_1_conv(v_286)
v_288 = self.pnnx_unique_86(v_287)
v_289 = self.model_23_one2one_cv4_0_2(v_288)
# v_290 = v_289.reshape(1, 1, 16384)
v_290 = v_289.reshape(1, 1, -1).transpose(1,2)

v_291 = self.model_23_one2one_cv4_1_0_conv(v_197)
v_292 = self.pnnx_unique_87(v_291)
v_293 = self.model_23_one2one_cv4_1_1_conv(v_292)
v_294 = self.pnnx_unique_88(v_293)
v_295 = self.model_23_one2one_cv4_1_2(v_294)
# v_296 = v_295.reshape(1, 1, 4096)
v_296 = v_295.reshape(1, 1, -1).transpose(1,2)

v_297 = self.model_23_one2one_cv4_2_0_conv(v_233)
v_298 = self.pnnx_unique_89(v_297)
v_299 = self.model_23_one2one_cv4_2_1_conv(v_298)
v_300 = self.pnnx_unique_90(v_299)
v_301 = self.model_23_one2one_cv4_2_2(v_300)
# v_302 = v_301.reshape(1, 1, 1024)
v_302 = v_301.reshape(1, 1, -1).transpose(1,2)

# ---合并三个尺度的 angle 预测 ---
v_303 = torch.cat((v_290, v_296, v_302), dim=1)

v_det = torch.cat((v_253, v_284), dim=2) 
return v_det, v_303