ncnn-android-yolov8

The YOLOv8 object detection

This is a sample ncnn android project, it depends on ncnn library and opencv

https://github.com/Tencent/ncnn

https://github.com/nihui/opencv-mobile

android apk file download

https://github.com/nihui/ncnn-android-yolov8/releases/latest

how to build and run

step1

https://github.com/Tencent/ncnn/releases

• Download ncnn-YYYYMMDD-android-vulkan.zip or build ncnn for android yourself
• Extract ncnn-YYYYMMDD-android-vulkan.zip into app/src/main/jni and change the ncnn_DIR path to yours in app/src/main/jni/CMakeLists.txt

step2

https://github.com/nihui/opencv-mobile

• Download opencv-mobile-XYZ-android.zip
• Extract opencv-mobile-XYZ-android.zip into app/src/main/jni and change the OpenCV_DIR path to yours in app/src/main/jni/CMakeLists.txt

step3

• Open this project with Android Studio, build it and enjoy!

some notes

• Android ndk camera is used for best efficiency
• Crash may happen on very old devices for lacking HAL3 camera interface
• All models are manually modified to accept dynamic input shape
• Most small models run slower on GPU than on CPU, this is common
• FPS may be lower in dark environment because of longer camera exposure time

screenshot

guidelines for converting YOLOv8 models

1. install

pip3 install -U ultralytics pnnx ncnn

2. export yolov8 torchscript

yolo export model=yolov8n.pt format=torchscript
yolo export model=yolov8n-seg.pt format=torchscript
yolo export model=yolov8n-pose.pt format=torchscript
yolo export model=yolov8n-cls.pt format=torchscript
yolo export model=yolov8n-obb.pt format=torchscript

3. convert torchscript with static shape

For classification models, step 1-3 is enough.

pnnx yolov8n.torchscript
pnnx yolov8n-seg.torchscript
pnnx yolov8n-pose.torchscript
pnnx yolov8n-cls.torchscript
pnnx yolov8n-obb.torchscript

4. modify pnnx model script for dynamic shape inference manually

Edit yolov8n_pnnx.py / yolov8n_seg_pnnx.py / yolov8n_pose_pnnx.py / yolov8n_obb_pnnx.py

• modify reshape to support dynamic image sizes
• permute tensor before concat and adjust concat axis (permutations are faster on smaller tensors)
• drop post-process part (we implement the post-process externally to avoid invalid bounding box coordinate calculations below the threshold, which is faster)

model	before	after
det	v_165 = v_142.view(1, 144, 6400) v_166 = v_153.view(1, 144, 1600) v_167 = v_164.view(1, 144, 400) v_168 = torch.cat((v_165, v_166, v_167), dim=2) # ...	v_165 = v_142.view(1, 144, -1).transpose(1, 2) v_166 = v_153.view(1, 144, -1).transpose(1, 2) v_167 = v_164.view(1, 144, -1).transpose(1, 2) v_168 = torch.cat((v_165, v_166, v_167), dim=1) return v_168
seg	v_144 = v_143.view(1, 32, 6400) v_150 = v_149.view(1, 32, 1600) v_156 = v_155.view(1, 32, 400) v_157 = torch.cat((v_144, v_150, v_156), dim=2) # ... v_191 = v_168.view(1, 144, 6400) v_192 = v_179.view(1, 144, 1600) v_193 = v_190.view(1, 144, 400) v_194 = torch.cat((v_191, v_192, v_193), dim=2) # ... v_215 = (v_214, v_138, ) return v_215	v_144 = v_143.view(1, 32, -1).transpose(1, 2) v_150 = v_149.view(1, 32, -1).transpose(1, 2) v_156 = v_155.view(1, 32, -1).transpose(1, 2) v_157 = torch.cat((v_144, v_150, v_156), dim=1) # ... v_191 = v_168.view(1, 144, -1).transpose(1, 2) v_192 = v_179.view(1, 144, -1).transpose(1, 2) v_193 = v_190.view(1, 144, -1).transpose(1, 2) v_194 = torch.cat((v_191, v_192, v_193), dim=1) return v_194, v_157, v_138
pose	v_137 = v_136.view(1, 51, 6400) v_143 = v_142.view(1, 51, 1600) v_149 = v_148.view(1, 51, 400) v_150 = torch.cat((v_137, v_143, v_149), dim=-1) # ... v_184 = v_161.view(1, 65, 6400) v_185 = v_172.view(1, 65, 1600) v_186 = v_183.view(1, 65, 400) v_187 = torch.cat((v_184, v_185, v_186), dim=2) # ...	v_137 = v_136.view(1, 51, -1).transpose(1, 2) v_143 = v_142.view(1, 51, -1).transpose(1, 2) v_149 = v_148.view(1, 51, -1).transpose(1, 2) v_150 = torch.cat((v_137, v_143, v_149), dim=1) # ... v_184 = v_161.view(1, 65, -1).transpose(1, 2) v_185 = v_172.view(1, 65, -1).transpose(1, 2) v_186 = v_183.view(1, 65, -1).transpose(1, 2) v_187 = torch.cat((v_184, v_185, v_186), dim=1) return v_187, v_150
obb	v_137 = v_136.view(1, 1, 16384) v_143 = v_142.view(1, 1, 4096) v_149 = v_148.view(1, 1, 1024) v_150 = torch.cat((v_137, v_143, v_149), dim=2) # ... v_186 = v_163.view(1, 79, 16384) v_187 = v_174.view(1, 79, 4096) v_188 = v_185.view(1, 79, 1024) v_189 = torch.cat((v_186, v_187, v_188), dim=2) # ...	v_137 = v_136.view(1, 1, -1).transpose(1, 2) v_143 = v_142.view(1, 1, -1).transpose(1, 2) v_149 = v_148.view(1, 1, -1).transpose(1, 2) v_150 = torch.cat((v_137, v_143, v_149), dim=1) # ... v_186 = v_163.view(1, 79, -1).transpose(1, 2) v_187 = v_174.view(1, 79, -1).transpose(1, 2) v_188 = v_185.view(1, 79, -1).transpose(1, 2) v_189 = torch.cat((v_186, v_187, v_188), dim=1) return v_189, v_150

5. re-export yolov8 torchscript

python3 -c 'import yolov8n_pnnx; yolov8n_pnnx.export_torchscript()'
python3 -c 'import yolov8n_seg_pnnx; yolov8n_seg_pnnx.export_torchscript()'
python3 -c 'import yolov8n_pose_pnnx; yolov8n_pose_pnnx.export_torchscript()'
python3 -c 'import yolov8n_obb_pnnx; yolov8n_obb_pnnx.export_torchscript()'

6. convert new torchscript with dynamic shape

Note the shape difference for obb model

pnnx yolov8n_pnnx.py.pt inputshape=[1,3,640,640] inputshape2=[1,3,320,320]
pnnx yolov8n_seg_pnnx.py.pt inputshape=[1,3,640,640] inputshape2=[1,3,320,320]
pnnx yolov8n_pose_pnnx.py.pt inputshape=[1,3,640,640] inputshape2=[1,3,320,320]
pnnx yolov8n_obb_pnnx.py.pt inputshape=[1,3,1024,1024] inputshape2=[1,3,512,512]

7. now you get ncnn model files

mv yolov8n_pnnx.py.ncnn.param yolov8n.ncnn.param
mv yolov8n_pnnx.py.ncnn.bin yolov8n.ncnn.bin
mv yolov8n_seg_pnnx.py.ncnn.param yolov8n_seg.ncnn.param
mv yolov8n_seg_pnnx.py.ncnn.bin yolov8n_seg.ncnn.bin
mv yolov8n_pose_pnnx.py.ncnn.param yolov8n_pose.ncnn.param
mv yolov8n_pose_pnnx.py.ncnn.bin yolov8n_pose.ncnn.bin
mv yolov8n_obb_pnnx.py.ncnn.param yolov8n_obb.ncnn.param
mv yolov8n_obb_pnnx.py.ncnn.bin yolov8n_obb.ncnn.bin