ORIS_AI

ORIS_AI is a deep learning framework specialized for on-device in the ORIS (On-device Robot Intelligence SW-System) project.

1. Supported Layers

• Input: OpenCV Mat
• Common: Concat, Convolution, Depthwise Convolution, ElementWise (SUM), MatMul, MaxPooling, Softmax. Split, Transpose
• Activation: SiLU
• Custom (for YOLO v8): C2f,BottleNeck (for C2f/C3k), DFL, SPPF, Proto
• Custom (for YOLO v11): C3k, C3k2, C2PSA, PSABlock, Attention
• Output (for YOLO v8/v11): Detect, Segment

2. Supported DNNs

• Detection: YOLO v8/v11
• Segmentation: YOLO v8/v11

3. Requirements

Component	Recommended version	Command to check the version
JetPack	6.2.1	dpkg -l | grep nvidia-jetpack
gcc/g++	11.4.0	gcc --version
cmake	≥ 3.22.1	cmake --version
cudatoolkit	12.6	nvcc --version
cudnn	9.3	cat {Your CUDNN Install Path}/cudnn_version.h
opencv	4.8	python3 -c "import cv2; print(cv2.version)"
protobuf	3.12.4	protoc --version

4. Prerequisites

• cmake [Required]
• eigen [Required]
• protobuf [Required]
• glog [Required]
• OpenCV [Required] : https://github.com/opencv/opencv
• OpenBLAS [Required] : https://github.com/xianyi/OpenBLAS.git

4-1. cmake

We recommend using the latest version of cmake. The latest version of cmake can be found at: [https://cmake.org/download/]

If needed, cmake can also be installed using the following commands. However, please note that installing via apt-get may not provide the latest version of cmake:

$ sudo apt-get install cmake
$ sudo apt-get install cmake-curses-gui # Only needed if you plan to use ccmake

4-2. eigen

$ sudo apt-get install libeigen3-dev

4-3. protobuf

$ sudo apt-get install libprotobuf-dev protobuf-compiler

4-4. glog

$ sudo apt-get install libgoogle-glog-dev

4-5. OpenBLAS

4-5-1. Build OpenBLAS

$ sudo apt-get install build-essential gfortran
$ git clone https://github.com/xianyi/OpenBLAS.git {OpenBLAS_Path}
$ cd {OpenBLAS_Path}
$ make FC=gfortran -j$(nproc)
$ make install PREFIX=/usr/local
$ ldconfig

4-5-2. Configure OPENBLAS_NUM_THREADS

$ cd ~
$ vi .bashrc

Then add OPENBLAS_NUM_THREADS=n to the last line. n is the maximum number of threads that your CPU supports. If you want to check the number of CPU cores in the current system, enter the command below.

$ grep -c processor /proc/cpuinfo

After editing the shell environment, reflect the edited shell environment and check the modified contents.

$ source .bashrc
$ echo $OPENBLAS_NUM_THREADS

4-6. OpenCV

📢 Notice on GPL License

OpenCV itself is released under the BSD 3-Clause License.
However, when building or installing OpenCV, additional external libraries may introduce GPL obligations.

• If GPL-related libraries such as libx264-dev or libxvidcore-dev are included, the resulting OpenCV build may be subject to the GPL license, requiring source code disclosure when redistributed.
• When enabling WITH_FFMPEG=ON, please check whether your FFmpeg installation is built in LGPL-only mode or whether it links to GPL codecs (e.g., x264).
• The option OPENCV_ENABLE_NONFREE=ON enables patented algorithms and should remain OFF unless you have verified that your use case allows it.

📌 Important

• For personal or research use, GPL obligations are generally not an issue.
• For distribution, commercial use, or providing binaries externally, you must verify whether GPL libraries are linked and, if so, ensure full compliance with the GPL license.
• If you want to avoid GPL obligations, build OpenCV without GPL-related libraries and rely only on BSD/LGPL components.

---

4-6-1. Install prerequisites

Install only the required dependencies for your environment. Avoid GPL-related packages if you want to distribute without GPL obligations.

4-6-2. Build OpenCV

Use the appropriate build options for your environment. By default, it is recommended to set WITH_FFMPEG=OFF and OPENCV_ENABLE_NONFREE=OFF.

5. How to compile ORIS_AI

The installation path of ORIS_AI (currently /ORIS_AI) needs to be modified to suit your environment.

5-1. Native compile

$ cd /ORIS_AI
$ mkdir build
$ cd build
$ cmake ..
$ make -j$(nproc)

5-2. Cmake configuration

Use ccmake

$ cd /ORIS_AI/build
$ ccmake ..

5.3. Configure the library path

Create symbolic links for the shared libraries to /usr/local/lib.

$ sudo ln -s /ORIS_AI/lib/aarch_64/libORISCore.so /usr/local/lib/libORISCore.so
$ sudo ln -s /ORIS_AI/lib/aarch_64/libORISCUDA.so /usr/local/lib/libORISCUDA.so

6. How to run ORIS_AI

The source codes of example are located in the following path.

/ORIS_AI/src/oris_ai/examples

The binaries of example are located in the following path.

/ORIS_AI/build/bin

7. Example

When using GPU, the ORIS AI inference engine is executed as an asynchronous stream, and when using CUDA, if you want to measure the actual execution time of each step, you must use the sync option.

7-1. YOLOv8 Official Model

7-1-1. YOLOv8n Detection

• /ORIS_AI/src/oris_ai/examples/test_yolo_v8.cc

Usage: ./test_yolo_v8 [OPTIONS]
Options:
  -c      Use CPU for inference. Default is GPU.
  -sync   Synchronize CUDA before timing. (For accurate GPU timing)
Example:
  ./test_yolo_v8 -c
  ./test_yolo_v8 -sync

7-1-2. YOLOv8n Segmentation

• /ORIS_AI/src/oris_ai/examples/test_yolo_v8_seg.cc

Usage: ./test_yolo_v8_seg [OPTIONS]
Options:
  -c      Use CPU for inference. Default is GPU.
  -sync   Synchronize CUDA before timing. (For accurate GPU timing)
Example:
  ./test_yolo_v8_seg -c
  ./test_yolo_v8_seg -sync

7-2. YOLOv11 Official Model

7-2-1. YOLOv11n Detection

• /ORIS_AI/src/oris_ai/examples/test_yolo_v11.cc

Usage: ./test_yolo_v11 [OPTIONS]
Options:
  -c      Use CPU for inference. Default is GPU.
  -sync   Synchronize CUDA before timing. (For accurate GPU timing)
Example:
  ./test_yolo_v11 -c
  ./test_yolo_v11 -sync

7-2-2. YOLOv11n Segmentation

• /ORIS_AI/src/oris_ai/examples/test_yolo_v11_seg.cc

Usage: ./test_yolo_v11_seg [OPTIONS]
Options:
  -c      Use CPU for inference. Default is GPU.
  -sync   Synchronize CUDA before timing. (For accurate GPU timing)
Example:
  ./test_yolo_v11_seg -c
  ./test_yolo_v11_seg -sync