📢 News

Install dependencies

Multiple versions of Python

# suppose Python 3.7.5 is installed in default pathexport LD_LIBRARY_PATH=/usr/local/python3.7.5/lib:$LD_LIBRARY_PATHexport PATH=/usr/local/python3.7.5/bin:$PATH

Install CANN

Build OpenCV with CANN

# replace ${cann_install_prefix} with your pathsource${cann_install_prefix}/ascend-toolkit/set_env.sh

git clone https://github.com/opencv/opencv.gitcd opencvmkdir buildcd buildcmake -D WITH_CANN=ON \      -D BUILD_opencv_gapi=OFF \      -D CMAKE_INSTALL_PREFIX=install ..# ensure you see 'CANN: YES' in the end of the log# Note: you could append "-j 8" in the following command for multi-job speedup.#       More jobs are used, more memory is needed.cmake --build. --target install

# replace the value of PYTHON3_EXECUTABLE to your path to python binary# replace the value of PYTHON3_LIBRARY to your path to python library (where you can find libpython3.x.so)# replace the value of PYTHON3_INCLUDE_DIR to your path to the python include directory (where you can find Python.h)cmake -D WITH_CANN=ON\      -D CMAKE_INSTALL_PREFIX=install \      -D BUILD_opencv_python2=OFF \      -D BUILD_opencv_python3=ON \      -D BUILD_opencv_gapi=OFF \      -D PYTHON3_EXECUTABLE=/usr/local/python3.7.5/bin/python3.7m \      -D PYTHON3_LIBRARY=/usr/local/python3.7.5/lib/libpython3.7m.so \      -D PYTHON3_INCLUDE_DIR=/usr/local/python3.7.5/include/python3.7m \      ..

Samples

Get models

Samples in Python

Samples in C++

OpenCV Zoo benchmark

Attachments

Python demos

importnumpyasnpimportcv2ascvdefpreprocess(image):out=image.copy()out=cv.resize(out, (256,256))out=out[16:240,16:240, :]out=cv.dnn.blobFromImage(out,1.0/255.0,mean=(0.485,0.456,0.406),swapRB=True)out=out/np.array([0.229,0.224,0.225]).reshape(1,-1,1,1)returnoutdefsoftmax(blob,axis=1):out=blob.copy().astype(np.float64)e_blob=np.exp(out)returne_blob/np.sum(e_blob,axis=axis)image=cv.imread("/path/to/image")# replace with the path to your imageinput_blob=preprocess(image)net=cv.dnn.readNet("/path/to/image_classification_mobilenetv1_2022apr.onnx")# replace with the path to the modelnet.setPreferableBackend(cv.dnn.DNN_BACKEND_CANN)net.setPreferableTarget(cv.dnn.DNN_TARGET_NPU)net.setInput(input_blob)out=net.forward()prob=softmax(out,axis=1)_,max_prob,_,max_loc=cv.minMaxLoc(prob)print("cls = {}, score = {:.4f}".format(max_loc[0],max_prob))

importnumpyasnpimportcv2ascvdefpreprocess(image):out=image.copy()out=cv.resize(out, (256,256))out=out[16:240,16:240, :]out=cv.dnn.blobFromImage(out,1.0/255.0,mean=(0.485,0.456,0.406),swapRB=True)out=out/np.array([0.229,0.224,0.225]).reshape(1,-1,1,1)returnoutdefsoftmax(blob,axis=1):out=blob.copy().astype(np.float64)e_blob=np.exp(out)returne_blob/np.sum(e_blob,axis=axis)image=cv.imread("/path/to/image")# replace with the path to your imageinput_blob=preprocess(image)net=cv.dnn.readNet("/path/to/image_classification_ppresnet50_2022jan.onnx")# # replace with the path to the modelnet.setPreferableBackend(cv.dnn.DNN_BACKEND_CANN)net.setPreferableTarget(cv.dnn.DNN_TARGET_NPU)net.setInput(input_blob)output_blob=net.forward("save_infer_model/scale_0.tmp_0")prob=softmax(output_blob,axis=1)_,max_prob,_,max_loc=cv.minMaxLoc(prob)print("cls = {}, score = {:.4f}".format(max_loc[0],max_prob))

importnumpyasnpimportcv2ascvdefpostprocess(blob,confidence_threshold=0.5,nms_threshold=0.5):out=blob.copy()strides= [8,16,32]hsizes= [80,40,20]wsizes= [80,40,20]grids= []expanded_strides= []forhsize,wsize,strideinzip(hsizes,wsizes,strides):xv,yv=np.meshgrid(np.arange(hsize),np.arange(wsize))grid=np.stack((xv,yv),2).reshape(1,-1,2)grids.append(grid)shape=grid.shape[:2]expanded_strides.append(np.full((*shape,1),stride))grids=np.concatenate(grids,1)expanded_strides=np.concatenate(expanded_strides,1)out[..., :2]= (out[..., :2]+grids)*expanded_stridesout[...,2:4]=np.exp(out[...,2:4])*expanded_strides# retrieve bboxesbboxes=out[0, :, :4]bboxes_xyxy=np.ones_like(bboxes)# (n, 4)bboxes_xyxy[:,0]=bboxes[:,0]-bboxes[:,2]/2.bboxes_xyxy[:,1]=bboxes[:,1]-bboxes[:,3]/2.bboxes_xyxy[:,2]=bboxes[:,0]+bboxes[:,2]/2.bboxes_xyxy[:,3]=bboxes[:,1]+bboxes[:,3]/2.# retrieve scoresscores=out[0, :,4:5]*out[0, :,5:]max_scores=np.amax(scores,axis=1)max_scores_idx=np.argmax(scores,axis=1)out=np.concatenate([bboxes_xyxy,max_scores[:,None],max_scores_idx[:,None]],axis=1)# batched-nmsmax_coord=bboxes_xyxy.max()offsets=max_scores_idx* (max_coord+1)bboxes_for_nms=bboxes_xyxy+offsets[:,None]keep=cv.dnn.NMSBoxes(bboxes_for_nms.tolist(),max_scores.tolist(),confidence_threshold,nms_threshold)final_out=out[keep]returnfinal_outimage=cv.imread("/path/to/image")# replace with the path to your imageinput_blob=cv.dnn.blobFromImage(image,size=(640,640),swapRB=True)net=cv.dnn.readNet("/path/to/object_detection_yolox_2022nov.onnx")# replace with the path to the modelnet.setPreferableBackend(cv.dnn.DNN_BACKEND_CANN)net.setPreferableTarget(cv.dnn.DNN_TARGET_NPU)net.setInput(input_blob)out=net.forward()dets=postprocess(out)fordetindets:bbox=det[0:4].astype(np.int32)score=det[4]clsid=det[5].astype(np.int32)print("bbox: {}, score: {:.4f}, clsid: {}".format(bbox,score,clsid))

C++ demos

cmake_minimum_required(VERSION 3.5.1)project(cann_demo)# OpenCVfind_package(OpenCV 4.6.0 REQUIRED)include_directories(${OpenCV_INCLUDE_DIRS})# PP-ResNet50add_executable(ppresnet50 ppresnet50.cpp)target_link_libraries(ppresnet50${OpenCV_LIBS})# MobileNetV1add_executable(mobilenetv1 mobilenetv1.cpp)target_link_libraries(mobilenetv1${OpenCV_LIBS})# YOLOXadd_executable(yolox yolox.cpp)target_link_libraries(yolox${OpenCV_LIBS})

#include<iostream>#include<vector>#include"opencv2/opencv.hpp"voidpreprocess(const cv::Mat& src, cv::Mat& dst){    src.convertTo(dst, CV_32FC3);cv::cvtColor(dst, dst, cv::COLOR_BGR2RGB);// center cropcv::resize(dst, dst,cv::Size(256,256));    cv::Rectroi(16,16,224,224);    dst =dst(roi);    dst =cv::dnn::blobFromImage(dst,1.0/255.0,cv::Size(),cv::Scalar(0.485,0.456,0.406));cv::divide(dst,cv::Scalar(0.229,0.224,0.225), dst);}voidsoftmax(const cv::Mat& src, cv::Mat& dst,int axis=1){usingnamespacecv::dnn;    LayerParams lp;    Net netSoftmax;    netSoftmax.addLayerToPrev("softmaxLayer","Softmax", lp);    netSoftmax.setPreferableBackend(DNN_BACKEND_OPENCV);    netSoftmax.setInput(src);    cv::Mat out = netSoftmax.forward();    out.copyTo(dst);}intmain(int argc,char** argv){usingnamespacecv;    Mat image =imread("/path/to/image");// replace with the path to your image    Mat input_blob;preprocess(image, input_blob);    dnn::Net net =dnn::readNet("/path/to/image_classification_mobilenetv1_2022apr.onnx");// replace with the path to the model    net.setPreferableBackend(dnn::DNN_BACKEND_CANN);    net.setPreferableTarget(dnn::DNN_TARGET_NPU);    net.setInput(input_blob);    Mat out = net.forward();    Mat prob;softmax(out, prob,1);double min_val, max_val;    Point min_loc, max_loc;minMaxLoc(prob, &min_val, &max_val, &min_loc, &max_loc);    std::cout <<cv::format("cls = %d, score = %.4f\n", max_loc.x, max_val);return0;}

#include<iostream>#include<vector>#include"opencv2/opencv.hpp"voidpreprocess(const cv::Mat& src, cv::Mat& dst){    src.convertTo(dst, CV_32FC3);cv::cvtColor(dst, dst, cv::COLOR_BGR2RGB);// center cropcv::resize(dst, dst,cv::Size(256,256));    cv::Rectroi(16,16,224,224);    dst =dst(roi);    dst =cv::dnn::blobFromImage(dst,1.0/255.0,cv::Size(),cv::Scalar(0.485,0.456,0.406));cv::divide(dst,cv::Scalar(0.229,0.224,0.225), dst);}voidsoftmax(const cv::Mat& src, cv::Mat& dst,int axis=1){usingnamespacecv::dnn;    LayerParams lp;    Net netSoftmax;    netSoftmax.addLayerToPrev("softmaxLayer","Softmax", lp);    netSoftmax.setPreferableBackend(DNN_BACKEND_OPENCV);    netSoftmax.setInput(src);    cv::Mat out = netSoftmax.forward();    out.copyTo(dst);}intmain(int argc,char** argv){usingnamespacecv;    Mat image =imread("/path/to/image");// replace with the path to your image    Mat input_blob;preprocess(image, input_blob);    dnn::Net net =dnn::readNet("/path/to/image_classification_ppresnet50_2022jan.onnx");// replace with the path to the model    net.setPreferableBackend(dnn::DNN_BACKEND_CANN);    net.setPreferableTarget(dnn::DNN_TARGET_NPU);    net.setInput(input_blob);    Mat out = net.forward("save_infer_model/scale_0.tmp_0");    Mat prob;softmax(out, prob,1);double min_val, max_val;    Point min_loc, max_loc;minMaxLoc(prob, &min_val, &max_val, &min_loc, &max_loc);    std::cout <<cv::format("cls = %d, score = %.4f\n", max_loc.x, max_val);return0;}

#include<iostream>#include<vector>#include"opencv2/opencv.hpp"usingnamespacecv;cv::Matpostprocess(const cv::Mat& blob,constfloat confidence_threshold =0.5,constfloat nms_threshold =0.5){    std::vector<int> strides{8,16,32};    std::vector<int> hsizes{80,40,20};    std::vector<int> wsizes{80,40,20};    std::vector<Point2f>grids(8400);    std::vector<float>expanded_strides(8400);int i, j, k, l =0, h, w;for (i =0; i < hsizes.size(); i++)    {        h = hsizes[i];        w = wsizes[i];for (j =0; j < h; j++)        {for (k =0; k < w; k++)            {                Point2f grid{float(k),float(j)};                grids[l] = grid;                expanded_strides[l] =float(strides[i]);                l++;            }        }    }constfloat* p_delta = (constfloat*)blob.data;    Mat outs;    Matout(1,6, CV_32FC1);for (i =0; i <8400; i++)    {        j = i *85;        Point2f grid = grids[i];float expanded_stride = expanded_strides[i];// retrieve objectness scorefloat objectness = p_delta[j +4];// retrieve class scoresfloat max_score = -1.f;float max_idx = -1.f;float this_score;for (k =5; k <85; k++)        {            this_score = p_delta[j + k] * objectness;if (this_score > max_score)            {                max_score = this_score;                max_idx = k -5;            }        }if (max_score <0.5)continue;        out.at<float>(0,4) = max_score;        out.at<float>(0,5) = max_idx;// retrieve bboxfloat cx = (p_delta[j] + grid.x) * expanded_stride;float cy = (p_delta[j +1] + grid.y) * expanded_stride;float width =std::exp(p_delta[j +2]) * expanded_stride;float height =std::exp(p_delta[j +3]) * expanded_stride;        out.at<float>(0,0) = cx - width /2;        out.at<float>(0,1) = cy - height /2;        out.at<float>(0,2) = cx + width /2;        out.at<float>(0,3) = cy + height /2;        outs.push_back(out);    }    Mat dets = outs;if (dets.rows >1)    {// batched nmsfloat max_coord = -1;for (i =0; i < dets.rows; i++)for (j =0; j <4; j++)if (max_coord < dets.at<float>(i, j))                    max_coord = dets.at<float>(i, j);        std::vector<Rect2i> boxes;        std::vector<float> scores;float offsets;for (i =0; i < dets.rows; i++)        {            offsets = dets.at<float>(i,5) * (max_coord +1);            boxes.push_back(Rect2i(int(dets.at<float>(i,0) + offsets),int(dets.at<float>(i,1) + offsets),int(dets.at<float>(i,2) + offsets),int(dets.at<float>(i,3) + offsets))            );            scores.push_back(dets.at<float>(i,4));        }        std::vector<int> keep;dnn::NMSBoxes(boxes, scores,0.5,0.5, keep);        Mat dets_after_nms;for (auto idx : keep)            dets_after_nms.push_back(dets.row(idx));        dets = dets_after_nms;    }return dets;}intmain(int argc,char** argv){    Mat image =imread("/path/to/image");// replace with the path to your image    Mat input_blob =dnn::blobFromImage(image,1.0f,cv::Size(640,640),cv::Scalar(),true);    dnn::Net net =dnn::readNet("/path/to/object_detection_yolox_2022nov.onnx");// replace with the path to the model    net.setPreferableBackend(dnn::DNN_BACKEND_CANN);    net.setPreferableTarget(dnn::DNN_TARGET_NPU);    net.setInput(input_blob);    Mat out = net.forward();    Mat dets =postprocess(out);for (int i =0; i < dets.rows; i++)    {int x1 =int(dets.at<float>(i,0));int y1 =int(dets.at<float>(i,1));int x2 =int(dets.at<float>(i,2));int y2 =int(dets.at<float>(i,3));float score = dets.at<float>(i,4);int cls =int(dets.at<float>(i,5));        std::cout <<cv::format("box [%d, %d, %d, %d], score %f, class %d\n", x1, y1, x2, y2, score, cls);    }return0;}