zgabi / yolo.net

Alturos.Yolo

A state of the art real-time object detection system for C# (Visual Studio). This project has CPU and GPU support, with GPU the detection works much faster. The primary goal of this project is an easy use of yolo, this package is available on nuget and you must only install two packages to start detection. In the background we are use the Windows Yolo version of AlexeyAB/darknet. Send an image path or the byte array to yolo and receive the position of the detected objects. Our project is meant to return the object-type and -position as processable data. This library supports YoloV3 and YoloV2 Pre-Trained Datasets

NuGet

Quick install Alturos.Yolo over NuGet

PM> install-package Alturos.Yolo (C# wrapper and C++ dlls 28MB)
PM> install-package Alturos.YoloV2TinyVocData (YOLOv2-tiny Pre-Trained Dataset 56MB)

Object Detection

Example code

Detect the type and the position of an image (Automatic configuration)

var configurationDetector = new ConfigurationDetector();
var config = configurationDetector.Detect();
using (var yoloWrapper = new YoloWrapper(config))
{
	var items = yoloWrapper.Detect(@"image.jpg");
	//items[0].Type -> "Person, Car, ..."
	//items[0].Confidence -> 0.0 (low) -> 1.0 (high)
	//items[0].X -> bounding box
	//items[0].Y -> bounding box
	//items[0].Width -> bounding box
	//items[0].Height -> bounding box
}

Detect the type and the position of an image (Manual configuration)

using (var yoloWrapper = new YoloWrapper("yolov2-tiny-voc.cfg", "yolov2-tiny-voc.weights", "voc.names"))
{
	var items = yoloWrapper.Detect(@"image.jpg");
	//items[0].Type -> "Person, Car, ..."
	//items[0].Confidence -> 0.0 (low) -> 1.0 (high)
	//items[0].X -> bounding box
	//items[0].Y -> bounding box
	//items[0].Width -> bounding box
	//items[0].Height -> bounding box
}

Performance

It is important to use GPU mode for fast object detection. It is also important not to instantiate the wrapper over and over again. A further optimization is to transfer the images as byte stream instead of passing a file path. GPU detection is usually 10 times faster!

System requirements

• .NET Framework 4.6.1 or .NET standard 2.0
• Microsoft Visual C++ Redistributable for Visual Studio 2015, 2017 und 2019 x64

GPU requirements (optional)

It is important to use the mentioned version 10.2

1. Install the latest Nvidia driver for your graphic device
2. Install Nvidia CUDA Toolkit 10.2 (must be installed add a hardware driver for cuda support)
3. Download Nvidia cuDNN v7.6.5 for CUDA 10.2
4. Copy the cudnn64_7.dll from the output directory of point 2. into the project folder.

Ampere architecture (RTX3000) requirements (optional)

It is important to use the mentioned version CUDA11.0

1. Install the latest Nvidia driver for your graphic device
2. Install Nvidia CUDA Toolkit 11.0 (must be installed add a hardware driver for cuda support)
3. Download Nvidia cuDNN v8.0.5 for CUDA 11.0
4. Copy the cudnn64_8.dll from the output directory of point 2. into the project folder
5. Copy opencv_world452.dll to output directory.

Build requirements

• Visual Studio 2019

Benchmark / Performance

Average processing speed of test images bird1.png, bird2.png, car1.png, motorbike1.png

CPU

Processor	YOLOv2-tiny	YOLOv3	yolo9000
Intel i7 3770	260 ms	2200 ms	-
Intel Xeon E5-1620 v3	207 ms	4327 ms	-
Intel Xeon E3-1240 v6	182 ms	3213 ms	-

GPU

Graphic card	Single precision	Memory	Slot	YOLOv2-tiny	YOLOv3	yolo9000
NVIDIA Quadro K420	300 GFLOPS	2 GB	Single	-	-	-
NVIDIA Quadro K620	768 GFLOPS	2 GB	Single	-	-	-
NVIDIA Quadro K1200	1151 GFLOPS	4 GB	Single	-	-	-
NVIDIA Quadro P400	599 GFLOPS	2 GB	Single	-	-	-
NVIDIA Quadro P600	1117 GFLOPS	2 GB	Single	-	-	-
NVIDIA Quadro P620	1386 GFLOPS	2 GB	Single	-	-	-
NVIDIA Quadro P1000	1862 GFLOPS	4 GB	Single	-	-	-
NVIDIA Quadro P2000	3011 GFLOPS	5 GB	Single	-	-	-
NVIDIA Quadro P4000	5304 GFLOPS	8 GB	Single	-	-	-
NVIDIA Quadro P5000	8873 GFLOPS	16 GB	Dual	-	-	-
NVIDIA GeForce GT 710	366 GFLOPS	2 GB	Single	-	-	-
NVIDIA GeForce GT 730	693 GFLOPS	2-4 GB	Single	-	-	-
NVIDIA GeForce GT 1030	1098 GFLOPS	2 GB	Single	40 ms	160 ms	-
NVIDIA GeForce GTX 1060	4372 GFLOPS	6 GB	Dual	25 ms	100 ms	-

Pre-Trained Dataset

A Pre-Trained Dataset contains the Informations about the recognizable objects. A higher Processing Resolution detects object also if they are smaller but this increases the processing time. The Alturos.YoloV2TinyVocData package is the same as YOLOv2-tiny. You can download the datasets manually or integrate them automatically into the code.

//The download takes some time depending on the internet connection.
var repository = new YoloPreTrainedDatasetRepository();
await repository.DownloadDatasetAsync("YOLOv3", ".");

Model	Processing Resolution	Cfg	Weights	Names
YOLOv3	608x608	yolov3.cfg	yolov3.weights	coco.names
YOLOv3-tiny	416x416	yolov3-tiny.cfg	yolov3-tiny.weights	coco.names
YOLOv2	608x608	yolov2.cfg	yolov2.weights	coco.names
YOLOv2-tiny	416x416	yolov2-tiny.cfg	yolov2-tiny.weights	voc.names
yolo9000	448x448	yolo9000.cfg	yolo9000.weights	9k.names

yolo9000 require a data directory with this two files coco9k.map and 9k.tree. Merge files with this command type xaa xab > yolo9000.weights

Troubleshooting

If you have some error like DllNotFoundException use Dependencies to check all references are available for yolo_cpp_dll_gpu.dll

If you have some error like NotSupportedException check if you use the latest Nvidia driver

Debugging Tool for Nvidia Gpu

Check graphic device usage "%PROGRAMFILES%\NVIDIA Corporation\NVSMI\nvidia-smi.exe"

Directory Structure

You should have this files in your program directory.

.
├── Alturos.Yolo.dll              # C# yolo wrapper
├── yolo_cpp_dll_cpu.dll      # yolo runtime for cpu
├── yolo_cpp_dll_gpu.dll      # yolo runtime for gpu
├── cudnn64_7.dll             # required by yolo_cpp_dll_gpu (optional only required for gpu processig)
├── opencv_world340.dll       # required by yolo_cpp_dll_xxx (process image as byte data detect_mat)
├── pthreadGC2.dll            # required by yolo_cpp_dll_xxx (POSIX Threads)
├── pthreadVC2.dll            # required by yolo_cpp_dll_xxx (POSIX Threads)
├── msvcr100.dll              # required by pthread (POSIX Threads)

Annotation Tool

To marking bounded boxes of objects in images for training neural network you can use

• Alturos.ImageAnnotation
• VoTT

Dataset of tagged images

• http://cocodataset.org/
• https://storage.googleapis.com/openimages/web/index.html

yolo.net's Issues

Performance and yoloWrapper ('Attempted to read or write protected memory)

I was looking ot see if I could improve the performance of some of the common examples. I posted in the original, but I realise this Fork is actually the one I am using from nuget.

One of the notes in the main doc page says " It is also important not to instantiate the wrapper over and over again.". Which is actually one of the first things I have tried. However, if you do not instantiate a new YoloWrapper every time before you do a .Detect with it, you get this error "System.AccessViolationException: 'Attempted to read or write protected memory. This is often an indication that another memory is corrupt.'"

So, it's required to do yoloWrapper = new YoloWrapper(config); every time, even though the docs are saying this is wasteful (which it is!)

Does anyone know how to make this work whilst re-using the already initialised YolowWrapper?

Guide for dummies

Hi!

As I got some good help from users here I decided try to help others coming after me. I got it to work for 30 and 10- series (Don't have a 20-card ut would expect it works for it) but there are challenges to get everything right. So for those of you out there needing some help!

1. Install the latest Nvidia drivers.
2. Install Nvidia CUDA toolkit 11.0 (The exact version is required. This might change so check the first page for the version)
3. Download and install Nvidia cuDNN. Check the first page for the exact version.
4. Copy the DLL files from the zip file in step 3 to your output directory. cudnn64_8.dll INCLUDING the other DLLs in the same directory.
5. Copy opencv_world452.dll (see first page for exact version) to the output directory.
6. Use the yolo_cpp_dll_gpu.dll in https://github.com/zgabi/Yolo.Net/tree/master/lib/Cuda11
7. User the yolo_cpp_dll_cpu.dll in https://github.com/zgabi/Yolo.Net/tree/master/lib/Cuda10

Make sure you have all the DLLs in the output directory:

• CuDNN files (cudnn64_8.dll + the extra in the same directory)
• opencv_world452.dll
• yolo_cpp_dll_gpu.dll
• yolo_cpp_dll_cpu.dll
• The other required DLLs (see first page) pthreadGC2.dll, pthreadVC2.dll , msvcr100.dll

Still don't work? Here are some ideas.

• For me I had to have the yolo_cpp_dll_gpu, yolo_cpp_dll_cpu and cudnn64_8.dll in both the output directory AND a sub directory called dll.

• Make sure you have the right yolo_cpp_dll_gpu.dll in the dll folder and output folder. It should be around 3MB. (it is the one in https://github.com/zgabi/Yolo.Net/tree/master/lib/Cuda11.0)

• As the wrapper check this file binary to look for the right version of CUDA, having the wrong version of yolo_cpp_dll_gpu will get you an error "CUDNN does not exist", but that is because it is looking for the wrong CUDNN. It fooled me once...

• Make sure you compile in x64.

where is your nuget package ?

Hi,

Cuda 11 and RTX3000

Is CUDA 11.2 and RTX3000 series supported?

Support yolo v4

Hi can you support yolo v4 in Yolo.Net

Now use segmentation?