-
Notifications
You must be signed in to change notification settings - Fork 6
/
demo.m
178 lines (163 loc) · 7.8 KB
/
demo.m
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
%% yolov5-matlab
% TheMatrix [email protected]
%
% Import yolov5*.onnx for inference, including yolov5s.onnx, yolov5m.onnx, yolov5l.onnx,
% yolov5x.onnx, the original output dimension is 1*255*H*W(Other dimension formats
% can be slightly modified), import (importONNXFunction) + detection in matlab
% Head decoding output.
%
% 导入yolov5*.onnx进行推理（inference），包括yolov5s.onnx,yolov5m.onnx,yolov5l.onnx,yolov5x.onnx,其原始输出维度为1*255*H*W(其他维度形式可稍作修改),
% 在matlab中导入(importONNXFunction)+检测头解码输出。
%
%
%% Requirements
% Matlab R2021a or higher,the newer the better,no other dependencies.
%% import model
model = "./onnxPretrainedModels/"+"yolov5s.onnx";
customYoloV5FcnName = 'yolov5fcn';
inputSize = [640,640];
throushHold = 0.3;
nmsThroushHold = 0.5;
outs = cell(3,1); % 3个检测head输出
classesNames = categorical(readlines("coco.names"));
colors = randi(255,length(classesNames),3);
params = importONNXFunction(model,customYoloV5FcnName);
%% 摄像头视频流识别
cap = webcam();
player = vision.DeployableVideoPlayer();
image = cap.snapshot(); % image = imread('images/person.jpg');
step(player, image);
[H,W,~] = size(image);
while player.isOpen()
image = cap.snapshot(); % image = imread('images/person.jpg');
img = imresize(image,inputSize);
img = rescale(img,0,1);% 转换到[0,1]
img = permute(img,[3,1,2]);
img = dlarray(reshape(img,[1,size(img)])); % n*c*h*w，[0,1],RGB顺序
if canUseGPU()
img = gpuArray(img);
end
t1 = tic;
[outs{:}] = feval(customYoloV5FcnName,img,params,...
'Training',false,...
'InputDataPermutation','none',...
'OutputDataPermutation','none');% or call the function directly
fprintf('yolov5预测耗时：%.2f 秒\n',toc(t1));% yolov5s大概0.25秒，yolov5m大概0.75秒，yolov5l大概1.49秒，yolov5x大概2.81秒
outFeatures = yolov5Decode(outs,H,W);
%% 阈值过滤+NMS处理
scores = outFeatures(:,5);
outFeatures = outFeatures(scores>throushHold,:);
allBBoxes = outFeatures(:,1:4);
[maxScores,indxs] = max(outFeatures(:,6:end),[],2);
allScores = maxScores;
allLabels = classesNames(indxs);
% NMS非极大值抑制
if ~isempty(allBBoxes)
[bboxes,scores,labels] = selectStrongestBboxMulticlass(allBBoxes,allScores,allLabels,...
'RatioType','Min','OverlapThreshold',nmsThroushHold);
annotations = string(labels) + ": " + string(scores);
[~,ids] = ismember(labels,classesNames);
color = colors(ids,:);
image = insertObjectAnnotation(image,...
'rectangle',bboxes,cellstr(annotations),...
'Color',color,...
'LineWidth',3);
end
step(player,image);
end
release(player);
%% support function
function outPutFeatures = yolov5Decode(featuremaps,oriHight,oriWidth,anchors)
% 功能：根据anchors对yolov5.onnx原始输出特征图进行解码，直接获得解码后的输出特征矩阵
% 输入：
% featuremaps： numHeads*1大小的cell，每个cell保存当前检测头的featuremap,为bs*[(4+1+nc)*na]*h*w大小特征图矩阵
% oriHight:原始输入图像的高
% oriWidth:原始输入图像的宽
% anchors： 对应模型yolov5.onnx的所有anchors,na*2大小，每行形如[width,height]顺序排列，na为anchors的数量
% 输出：
% outPutFeatures： M*(5+nc)或者bs*M*(5+nc)大小
% ,为bs*M*[x,y,w,h,Pobj,p1,p2,...,pn]大小的形式矩阵，如果是单张图像检测，则输出大小为M*(5+nc)，否则是bs*M*(5+nc),
% 其中，M为检测框的数量，bs为图片数量，nc为训练网络dlnet类别数量，x,y,w,h分别是输入图片上对应的x,y,width,height，Pobj
% 为物体概率，p1,p2,...,pn分别为对应coco.names类别概率
% 注意：
% 1、假设yolov5.onnx输入图像大小为1*3*640*640,[0,1]范围float，RGB顺序；
% 2、输出有3个head,每个head大小形如为1*255*gh*gw，其中gh,gw分别表示特征图高和宽，255表示na*[cx,cy,w,h,Pobj,p1,p2,...pn]顺序含义；
% 3、小值的anchor对应原始输出的大特征图，大值的anchor对应原始输出的小特征图，anchors“从小到大”顺序传入，3个anchor为一组；
% 4、每一个检测head对应3个anchor
% 5、检测目标类别数量为coco中的80
% 其他onnx模型格式类推，大同小异
%
% reference:
% 1、https://github.com/cuixing158/yolov3-yolov4-matlab/blob/master/utils/yolov3v4Predict.m
% 2、https://github.com/ultralytics/yolov5/blob/master/models/yolo.py
% 3、https://github.com/onnx/models
%
% author: cuixingxing
% email:[email protected]
% 2021.3.17创建
%
arguments
featuremaps (:,1) cell
oriHight (1,1) double
oriWidth (1,1) double
anchors (:,2) double = [10,13; 16,30; 33,23;...
30,61; 62,46; 59,119;...
116,90; 156,198; 373,326]
end
%% yolov5*.onnx known params
inputSize = 640;%输入网络图像大小，正方形图像输入
na = 3;% 每个检测head对应anchor的数量
nc = 80; % coco类别数量
%% decode
scaledX = inputSize./oriWidth;
scaledY = inputSize./oriHight;
outPutFeatures = [];
numberFeaturemaps = length(featuremaps);
for i = 1:numberFeaturemaps
currentFeatureMap = featuremaps{i};% bs*[(4+1+nc)*na]*h*w大小
currentAnchors = anchors(na*(i-1)+1:na*i,:); % na*2
numY = size(currentFeatureMap,3);
numX = size(currentFeatureMap,4);
stride = inputSize./numX;
% reshape currentFeatureMap到有意义的维度，bs*[(4+1+nc)*na]*h*w --> h*w*(5+nc)*na*bs
% --> bs*na*h*w*(5+nc),最终的维度方式与yolov5官网兼容
bs = size(currentFeatureMap,1);
h = numY;
w = numX;
currentFeatureMap = reshape(currentFeatureMap,bs,5+nc,na,h,w);% bs*(5+nc)*na*h*w
currentFeatureMap = permute(currentFeatureMap,[1,3,4,5,2]);% bs*na*h*w*(5+nc)
[~,~,yv,xv] = ndgrid(1:bs,1:na,0:h-1,0:w-1);% yv,xv大小都为bs*na*h*w，注意顺序，后面做加法维度标签要对应
gridXY = cat(5,xv,yv);% 第5维上扩展，大小为bs*na*h*w*2, x,y从1开始的索引
currentFeatureMap = sigmoid(currentFeatureMap); % yolov5是对所有值进行归一化，与yolov3/v4不同
currentFeatureMap(:,:,:,:,1:2) = (2*currentFeatureMap(:,:,:,:,1:2)-0.5 + gridXY).*stride; % 大小为bs*na*h*w*2,预测对应xy
anchor_grid = reshape(currentAnchors,1,na,1,1,2);% 此处anchor_grid大小为1*na*1*1*2，方便下面相乘
currentFeatureMap(:,:,:,:,3:4) = (currentFeatureMap(:,:,:,:,3:4)*2).^2.*anchor_grid;% 大小为bs*na*h*w*2
if nc == 1
currentFeatureMap(:,:,:,:,6) = 1;
end
currentFeatureMap = reshape(currentFeatureMap,bs,[],5+nc);% bs*N*(5+nc)
if isempty(outPutFeatures)
outPutFeatures = currentFeatureMap;
else
outPutFeatures = cat(2,outPutFeatures,currentFeatureMap);% bs*M*(5+nc)
end
end
%% 坐标转换到原始图像上
% [cx,cy,w,h]，yolov5.onnx基准图像大小（1*3*640*640）----> [x,y,w,h],坐标基于原始图像大小（1*3*oriHight*oriWidth）
outPutFeatures = extractdata(outPutFeatures);% bs*M*(5+nc) ,为[x_center,y_center,w,h,Pobj,p1,p2,...,pn]
outPutFeatures(:,:,[1,3]) = outPutFeatures(:,:,[1,3])./scaledX;% x_center,width
outPutFeatures(:,:,[2,4]) = outPutFeatures(:,:,[2,4])./scaledY;% y_center,height
outPutFeatures(:,:,1) = outPutFeatures(:,:,1) -outPutFeatures(:,:,3)/2;% x
outPutFeatures(:,:,2) = outPutFeatures(:,:,2) -outPutFeatures(:,:,4)/2; % y
outPutFeatures = squeeze(outPutFeatures); % 如果是单张图像检测，则输出大小为M*(5+nc)，否则是bs*M*(5+nc)
if(canUseGPU())
outPutFeatures = gather(outPutFeatures); % 推送到CPU上
end
end
%% Reference:
% [1] <https://github.com/cuixing158/yolov3-yolov4-matlab/blob/master/utils/yolov3v4Predict.m
% yolov3-yolov4-matlab>
%
% [2] <https://github.com/ultralytics/yolov5/blob/master/models/yolo.py yolov5>
%
% [3] <https://github.com/onnx/models onnxModels>