This is a sample code showing object detection using Arm NN in two different modes:
The compiled application can take
as input and
with detections shown in bounding boxes, class labels and confidence.
This example utilizes OpenCV functions to capture and output video data.
This example was created on Ubuntu 20.04 with gcc and g++ version 9. If encountered any compiler errors while running with a different compiler version, you can install version 9 with:
sudo apt install gcc-9 g++-9
and add to every cmake command those compiler flags: -DCMAKE_C_COMPILER=gcc-9 -DCMAKE_CXX_COMPILER=g++-9
Object detection example build system does not trigger Arm NN compilation. Thus, before building the application, please ensure that Arm NN libraries and header files are available on your build platform. The application executable binary dynamically links with the following Arm NN libraries:
Pre compiled Arm NN libraries can be downloaded from https://github.com/ARM-software/armnn/releases/download/v21.11/ArmNN-linux-aarch64.tar.gz the "lib" and "include" directories should be taken together.
The build script searches for available Arm NN libraries in the following order:
Arm NN header files will be searched in parent directory of found libraries files under include directory, i.e. libraries found in /usr/lib or /usr/lib64 and header files in /usr/include (or ${ARMNN_LIB_DIR}/include).
Please see find_armnn.cmake for implementation details.
This application uses OpenCV (Open Source Computer Vision Library) for video stream processing. Your host platform may have OpenCV available through linux package manager. If this is the case, please install it using standard way.
sudo apt install python3-opencv
If not, our build system has a script to download and cross-compile required OpenCV modules as well as FFMPEG and x264 encoder libraries. The latter will build limited OpenCV functionality and application will support only video file input and video file output way of working. Displaying video frames in a window requires building OpenCV with GTK and OpenGL support.
The application executable binary dynamically links with the following OpenCV libraries:
and transitively depends on:
The application searches for above libraries in the following order:
If no OpenCV libraries were found, the cross-compilation build is extended with x264, ffmpeg and OpenCV compilation steps.
Note: Native build does not add third party libraries to compilation.
Please see find_opencv.cmake for implementation details.
This application uses Tensorflow Lite) version 2.5.0 for demonstrating use of 'armnnDelegate'. armnnDelegate is a library for accelerating certain TensorFlow Lite operators on Arm hardware by providing the TensorFlow Lite interpreter with an alternative implementation of the operators via its delegation mechanism. You may clone and build Tensorflow lite and provide the path to its root and output library directories through the cmake flags TENSORFLOW_ROOT and TFLITE_LIB_ROOT respectively. For implementation details see the scripts FindTfLite.cmake and FindTfLiteSrc.cmake
The application links with the Tensorflow lite library libtensorflow-lite.a
Example for Tensorflow Lite native compilation
sudo apt install build-essential git clone https://github.com/tensorflow/tensorflow.git cd tensorflow/tensorflow git checkout tags/v2.5.0 mkdir build && cd build cmake ../lite -DTFLITE_ENABLE_XNNPACK=OFF make
This application uses Flatbuffers) version 1.12.0 for serialization You may clone and build Flatbuffers and provide the path to its root directory through the cmake flag FLATBUFFERS_ROOT. Please see [FindFlatbuffers.cmake] for implementation details.
The application links with the Flatbuffers library libflatbuffers.a
Example for flatbuffer native compilation
wget -O flatbuffers-1.12.0.zip https://github.com/google/flatbuffers/archive/v1.12.0.zip unzip -d . flatbuffers-1.12.0.zip cd flatbuffers-1.12.0 mkdir install && cd install cmake .. -DCMAKE_INSTALL_PREFIX:PATH=`pwd` make install
There are two flows for building this application:
CMAKE_TOOLCHAIN_FILE - choose one of the available cross-compilation toolchain files:
cmake/aarch64-toolchain.cmakecmake/arm-linux-gnueabihf-toolchain.cmakeARMNN_LIB_DIR - point to the custom location of the Arm NN libs and headers.
OPENCV_LIB_DIR - point to the custom location of the OpenCV libs and headers.
BUILD_UNIT_TESTS - set to 1 to build tests. Additionally to the main application, object_detection_example-tests unit tests executable will be created.
For the Delegate file mode:
USE_ARMNN_DELEGATE - set to True to build the application with Tflite and delegate file mode. default is False.
TFLITE_LIB_ROOT - point to the custom location of Tflite lib
TENSORFLOW_ROOT - point to the custom location of Tensorflow root directory
FLATBUFFERS_ROOT - point to the custom location of Flatbuffers root directory
To build this application on a host platform, firstly ensure that required dependencies are installed: For example, for raspberry PI:
sudo apt-get update sudo apt-get -yq install pkg-config sudo apt-get -yq install libgtk2.0-dev zlib1g-dev libjpeg-dev libpng-dev libxvidcore-dev libx264-dev sudo apt-get -yq install libavcodec-dev libavformat-dev libswscale-dev ocl-icd-opencl-dev
To build demo application, create a build directory:
mkdir build
cd build
If you have already installed Arm NN and OpenCV:
Inside build directory, run cmake and make commands:
cmake .. make
This will build the following in bin directory:
If you have custom Arm NN and OpenCV location, use OPENCV_LIB_DIR and ARMNN_LIB_DIR options:
cmake -DARMNN_LIB_DIR=/path/to/armnn -DOPENCV_LIB_DIR=/path/to/opencv .. make
If you have build with Delegate file mode and have custom Arm NN, Tflite, and Flatbuffers locations, use the USE_ARMNN_DELEGATE flag together with TFLITE_LIB_ROOT, TENSORFLOW_ROOT, FLATBUFFERS_ROOT and ARMNN_LIB_DIR options:
cmake -DARMNN_LIB_DIR=/path/to/armnn/build/lib/ -DUSE_ARMNN_DELEGATE=True -DTFLITE_LIB_ROOT=/path/to/tensorflow/ -DTENSORFLOW_ROOT=/path/to/tensorflow/ -DFLATBUFFERS_ROOT=/path/to/flatbuffers/ .. make
This section will explain how to cross-compile the application and dependencies on a Linux x86 machine for arm host platforms.
You will require working cross-compilation toolchain supported by your host platform. For raspberry Pi 3 and 4 with glibc runtime version 2.28, the following toolchains were successfully used:
Choose aarch64-linux-gnu if lscpu command shows architecture as aarch64 or arm-linux-gnueabihf if detected architecture is armv71.
You can check runtime version on your host platform by running:
ldd --version
On build machine, install C and C++ cross compiler toolchains and add them to the PATH variable.
Install package dependencies:
sudo apt-get update sudo apt-get -yq install pkg-config
Package config is required by OpenCV build to discover FFMPEG libs.
To build demo application, create a build directory:
mkdir build
cd build
Inside build directory, run cmake and make commands:
Arm 32bit
cmake -DARMNN_LIB_DIR=<path-to-armnn-libs> -DCMAKE_TOOLCHAIN_FILE=cmake/arm-linux-gnueabihf-toolchain.cmake .. make
Arm 64bit
cmake -DARMNN_LIB_DIR=<path-to-armnn-libs> -DCMAKE_TOOLCHAIN_FILE=cmake/aarch64-toolchain.cmake .. make
Add -j flag to the make command to run compilation in multiple threads.
From the build directory, copy the following to the host platform:
The full list of libs after cross-compilation to copy on your board:
libarmnn.so libarmnn.so.31 libarmnn.so.31.0 For Arm NN public C++ API mode: libarmnnTfLiteParser.so libarmnnTfLiteParser.so.24.4 end For Delegate file mode: libarmnnDelegate.so libarmnnDelegate.so.25 libarmnnDelegate.so.25.0 libtensorflow-lite.a libflatbuffers.a end libavcodec.so libavcodec.so.58 libavcodec.so.58.54.100 libavdevice.so libavdevice.so.58 libavdevice.so.58.8.100 libavfilter.so libavfilter.so.7 libavfilter.so.7.57.100 libavformat.so libavformat.so.58 libavformat.so.58.29.100 libavutil.so libavutil.so.56 libavutil.so.56.31.100 libopencv_core.so libopencv_core.so.4.0 libopencv_core.so.4.0.0 libopencv_highgui.so libopencv_highgui.so.4.0 libopencv_highgui.so.4.0.0 libopencv_imgcodecs.so libopencv_imgcodecs.so.4.0 libopencv_imgcodecs.so.4.0.0 libopencv_imgproc.so libopencv_imgproc.so.4.0 libopencv_imgproc.so.4.0.0 libopencv_video.so libopencv_video.so.4.0 libopencv_video.so.4.0.0 libopencv_videoio.so libopencv_videoio.so.4.0 libopencv_videoio.so.4.0.0 libpostproc.so libpostproc.so.55 libpostproc.so.55.5.100 libswresample.a libswresample.so libswresample.so.3 libswresample.so.3.5.100 libswscale.so libswscale.so.5 libswscale.so.5.5.100 libx264.so libx264.so.160
Once the application executable is built, it can be executed with the following options:
To run object detection on a supplied video file and output result to a video file:
LD_LIBRARY_PATH=/path/to/armnn/libs:/path/to/opencv/libs ./object_detection_example --label-path /path/to/labels/file --video-file-path /path/to/video/file --model-file-path /path/to/model/file --model-name [YOLO_V3_TINY | SSD_MOBILE] --output-video-file-path /path/to/output/file
To run object detection on a supplied video file and output result to a window gui:
LD_LIBRARY_PATH=/path/to/armnn/libs:/path/to/opencv/libs ./object_detection_example --label-path /path/to/labels/file --video-file-path /path/to/video/file --model-file-path /path/to/model/file --model-name [YOLO_V3_TINY | SSD_MOBILE]
This application has been verified to work against the MobileNet SSD and the YOLO V3 tiny models, which can be downloaded along with their label sets from the Arm Model Zoo:
This section provides a walkthrough of the application, explaining in detail the steps:
After parsing user arguments, the chosen video file or stream is loaded into an OpenCV cv::VideoCapture object. We use IFrameReader interface and OpenCV specific implementation CvVideoFrameReader in our main function to capture frames from the source using the ReadFrame() function.
The CvVideoFrameReader object also tells us information about the input video. Using this information and application arguments, we create one of the implementations of the IFrameOutput interface: CvVideoFileWriter or CvWindowOutput. This object will be used at the end of every loop to write the processed frame to an output video file or gui window. CvVideoFileWriter uses cv::VideoWriter with ffmpeg backend. CvWindowOutput makes use of cv::imshow() function.
See GetFrameSourceAndSink function in Main.cpp for more details.
In order to interpret the result of running inference on the loaded network, it is required to load the labels associated with the model. In the provided example code, the AssignColourToLabel function creates a vector of pairs label - colour that is ordered according to object class index at the output node of the model. Labels are assigned with a randomly generated RGB color. This ensures that each class has a unique color which will prove helpful when plotting the bounding boxes of various detected objects in a frame.
Depending on the model being used, CreatePipeline function returns specific implementation of the object detection pipeline.
All operations with Arm NN and networks are encapsulated in ArmnnNetworkExecutor class.
The first step with Arm NN SDK is to import a graph from file by using the appropriate parser.
The Arm NN SDK provides parsers for reading graphs from a variety of model formats. In our application we specifically focus on .tflite, .pb, .onnx models.
Based on the extension of the provided model file, the corresponding parser is created and the network file loaded with CreateNetworkFromBinaryFile() method. The parser will handle the creation of the underlying Arm NN graph.
Current example accepts tflite format model files, we use ITfLiteParser:
#include "armnnTfLiteParser/ITfLiteParser.hpp" armnnTfLiteParser::ITfLiteParserPtr parser = armnnTfLiteParser::ITfLiteParser::Create(); armnn::INetworkPtr network = parser->CreateNetworkFromBinaryFile(modelPath.c_str());
Arm NN supports optimized execution on multiple CPU and GPU devices. Prior to executing a graph, we must select the appropriate device context. We do this by creating a runtime context with default options with IRuntime().
For example:
#include "armnn/ArmNN.hpp" auto runtime = armnn::IRuntime::Create(armnn::IRuntime::CreationOptions());
We can optimize the imported graph by specifying a list of backends in order of preference and implement backend-specific optimizations. The backends are identified by a string unique to the backend, for example CpuAcc, GpuAcc, CpuRef.
For example:
std::vector<armnn::BackendId> backends{"CpuAcc", "GpuAcc", "CpuRef"};
Internally and transparently, Arm NN splits the graph into subgraph based on backends, it calls a optimize subgraphs function on each of them and, if possible, substitutes the corresponding subgraph in the original graph with its optimized version.
Using the Optimize() function we optimize the graph for inference and load the optimized network onto the compute device with LoadNetwork(). This function creates the backend-specific workloads for the layers and a backend specific workload factory which is called to create the workloads.
For example:
armnn::IOptimizedNetworkPtr optNet = Optimize(*network, backends, m_Runtime->GetDeviceSpec(), armnn::OptimizerOptions()); std::string errorMessage; runtime->LoadNetwork(0, std::move(optNet), errorMessage)); std::cerr << errorMessage << std::endl;
Parsers can also be used to extract the input information for the network. By calling GetSubgraphInputTensorNames we extract all the input names and, with GetNetworkInputBindingInfo, bind the input points of the graph. For example:
std::vector<std::string> inputNames = parser->GetSubgraphInputTensorNames(0); auto inputBindingInfo = parser->GetNetworkInputBindingInfo(0, inputNames[0]);
The input binding information contains all the essential information about the input. It is a tuple consisting of integer identifiers for bindable layers (inputs, outputs) and the tensor info (data type, quantization information, number of dimensions, total number of elements).
Similarly, we can get the output binding information for an output layer by using the parser to retrieve output tensor names and calling GetNetworkOutputBindingInfo().
All operations with Tflite and networks are encapsulated in ArmnnNetworkExecutor class.
The first step with Tflite is to build a model from file by using Flatbuffer model class. with that model we create the Tflite Interpreter.
#include <tensorflow/lite/interpreter.h> armnnTfLiteParser::ITfLiteParserPtr parser = armnnTfLiteParser::ITfLiteParser::Create();m_model = tflite::FlatBufferModel::BuildFromFile(modelPath.c_str()); tflite::ops::builtin::BuiltinOpResolver resolver; tflite::InterpreterBuilder(*m_model, resolver)(&m_interpreter);
after the Interpreter is created we allocate tensors using the AllocateTensors function of the Interpreter
m_interpreter->AllocateTensors();
Arm NN Delegate file is created using the ArmnnDelegate constructor The constructor accepts a DelegateOptions object that is created from the list of the preferred backends that we want to use, and the optimizerOptions object (optional). In this example we enable fast math and reduce all float32 operators to float16 optimizations. These optimizations can sometime improve the performance but can also cause degredation, depending on the model and the backends involved, therefore one should try it out and decide whether to use it or not.
#include <armnn_delegate.hpp> #include <DelegateOptions.hpp> #include <DelegateUtils.hpp> /* enable fast math optimization */ armnn::BackendOptions modelOptionGpu("GpuAcc", {{"FastMathEnabled", true}}); optimizerOptions.m_ModelOptions.push_back(modelOptionGpu); armnn::BackendOptions modelOptionCpu("CpuAcc", {{"FastMathEnabled", true}}); optimizerOptions.m_ModelOptions.push_back(modelOptionCpu); /* enable reduce float32 to float16 optimization */ optimizerOptions.m_ReduceFp32ToFp16 = true; armnnDelegate::DelegateOptions delegateOptions(preferredBackends, optimizerOptions); /* create delegate object */ std::unique_ptr<TfLiteDelegate, decltype(&armnnDelegate::TfLiteArmnnDelegateDelete)> theArmnnDelegate(armnnDelegate::TfLiteArmnnDelegateCreate(delegateOptions), armnnDelegate::TfLiteArmnnDelegateDelete);
Registering the Arm NN delegate file will provide the TensorFlow Lite interpreter with an alternative implementation of the operators that can be accelerated by the Arm hardware For example:
/* Register the delegate file */ m_interpreter->ModifyGraphWithDelegate(std::move(theArmnnDelegate));
Generic object detection pipeline has 3 steps, to perform data pre-processing, run inference and decode inference results in the post-processing step.
See ObjDetectionPipeline and implementations for MobileNetSSDv1 and YoloV3Tiny for more details.
Each frame captured from source is read as an cv::Mat in BGR format but channels are swapped to RGB in a frame reader code.
cv::Mat processed; ... objectDetectionPipeline->PreProcessing(frame, processed);
A pre-processing step consists of resizing the frame to the required resolution, padding and doing data type conversion to match the model input layer. For example, SSD MobileNet V1 that is used in our example takes for input a tensor with shape [1, 300, 300, 3] and data type uint8.
Pre-processing step returns cv::Mat object containing data ready for inference.
od::InferenceResults results; ... objectDetectionPipeline->Inference(processed, results);
Inference step will call ArmnnNetworkExecutor::Run method that will prepare input tensors and execute inference. We have two separate implementations of the ArmnnNetworkExecutor class and its functions including ArmnnNetworkExecutor::Run The first Implementation ArmnnNetworkExecutoris utilizing Arm NN C++ API, while the second implementation ArmnnNetworkExecutor is utilizing Tensorflow lite and its Delegate file mechanism.
A compute device performs inference for the loaded network using the EnqueueWorkload() function of the runtime context. For example:
//const void* inputData = ...; //outputTensors were pre-allocated before armnn::InputTensors inputTensors = {{ inputBindingInfo.first,armnn::ConstTensor(inputBindingInfo.second, inputData)}}; runtime->EnqueueWorkload(0, inputTensors, outputTensors);
We allocate memory for output data once and map it to output tensor objects. After successful inference, we read data from the pre-allocated output data buffer. See ArmnnNetworkExecutor::ArmnnNetworkExecutor and ArmnnNetworkExecutor::Run for more details.
Inside the PrepareTensors(..) function, the input frame is copied to the Tflite Interpreter input tensor, than the Tflite Interpreter performs inference for the loaded network using the Invoke() function. For example:
PrepareTensors(inputData, dataBytes); if (m_interpreter->Invoke() == kTfLiteOk)
After successful inference, we read data from the Tflite Interpreter output tensor and copy it to the outResults vector. See ArmnnNetworkExecutor::Run for more details.
The output from inference must be decoded to obtain information about detected objects in the frame. In the examples there are implementations for two networks but you may also implement your own network decoding solution here.
For SSD MobileNet V1 models, we decode the results to obtain the bounding box positions, classification index, confidence and number of detections in the input frame. See SSDResultDecoder for more details.
For YOLO V3 Tiny models, we decode the output and perform non-maximum suppression to filter out any weak detections below a confidence threshold and any redundant bounding boxes above an intersection-over-union threshold. See YoloResultDecoder for more details.
It is encouraged to experiment with threshold values for confidence and intersection-over-union (IoU) to achieve the best visual results.
The detection results are always returned as a vector of DetectedObject, with the box positions list containing bounding box coordinates in the form [x_min, y_min, x_max, y_max].
Post-processing step accepts a callback function to be invoked when the decoding is finished. We will use it to draw detections on the initial frame. With the obtained detections and using AddInferenceOutputToFrame function, we are able to draw bounding boxes around detected objects and add the associated label and confidence score.
//results - inference output objectDetectionPipeline->PostProcessing(results, [&frame, &labels](od::DetectedObjects detects) -> void { AddInferenceOutputToFrame(detects, *frame, labels); });
The processed frames are written to a file or displayed in a separate window.