【树莓派5 测评】 + 16.yolo-v3测试（zmj）
【树莓派5 测评】 + 16.yolo-v3测试（zmj）YOLO是You Look once的简称，是目标检测算法中比较常用的一种算法。YOLO-V3是目标检测算法中的YOLO算法的第三个版本。在这个版本中其实并没有太多的创新点，更多的是借鉴了前两个版本，但是却在保持速度的同时，在精度上做了优化。YOLO-V3所使用的主干特征提取网络为Darknet53，它作为一个单独的神经网络作用在图像上，将图像划分成多个区域并且预测边界框和每个区域的概率。//------darknet
https://pjreddie.com/darknet/
1. 安装YOLO安装YOLO的过程十分简单，下载darknet后编译即可。//------安装指令
//---下载并进入darknet目录
git clone https://github.com/pjreddie/darknet
cd darknet
//---编译安装
make
//------下载权重模型并测试
//---yolov3-tiny
wget https://pjreddie.com/media/files/yolov3-tiny.weights
./darknet detect cfg/yolov3-tiny.cfg yolov3-tiny.weights data/dog.jpg
//---yolov3
wget https://pjreddie.com/media/files/yolov3.weights
./darknet detector test cfg/coco.data cfg/yolov3.cfg yolov3.weights data/dog.jpg
2. 模型下载及测试测试yolov3-tiny和yolov3这两种模型。2.1 下载测试//------下载权重模型并测试
//---yolov3-tiny
wget https://pjreddie.com/media/files/yolov3-tiny.weights
./darknet detect cfg/yolov3-tiny.cfg yolov3-tiny.weights data/dog.jpg
//---yolov3
wget https://pjreddie.com/media/files/yolov3.weights
./darknet detector test cfg/coco.data cfg/yolov3.cfg yolov3.weights data/dog.jpg
//------Log信息
zhaomeijing@raspberrypi5:~/workspace/04_yolo/darknet$ ./darknet detect cfg/yolov3-tiny.cfg yolov3-tiny.weights data/dog.jpg
layer filters size input output
0 conv 163 x 3 / 1 416 x 416 x 3 -> 416 x 416 x160.150 BFLOPs
1 max 2 x 2 / 2 416 x 416 x16 -> 208 x 208 x16
2 conv 323 x 3 / 1 208 x 208 x16 -> 208 x 208 x320.399 BFLOPs
3 max 2 x 2 / 2 208 x 208 x32 -> 104 x 104 x32
4 conv 643 x 3 / 1 104 x 104 x32 -> 104 x 104 x640.399 BFLOPs
5 max 2 x 2 / 2 104 x 104 x64 -> 52 x52 x64
6 conv 1283 x 3 / 1 52 x52 x64 -> 52 x52 x 1280.399 BFLOPs
7 max 2 x 2 / 2 52 x52 x 128 -> 26 x26 x 128
8 conv 2563 x 3 / 1 26 x26 x 128 -> 26 x26 x 2560.399 BFLOPs
9 max 2 x 2 / 2 26 x26 x 256 -> 13 x13 x 256
10 conv 5123 x 3 / 1 13 x13 x 256 -> 13 x13 x 5120.399 BFLOPs
11 max 2 x 2 / 1 13 x13 x 512 -> 13 x13 x 512
12 conv 10243 x 3 / 1 13 x13 x 512 -> 13 x13 x10241.595 BFLOPs
13 conv 2561 x 1 / 1 13 x13 x1024 -> 13 x13 x 2560.089 BFLOPs
14 conv 5123 x 3 / 1 13 x13 x 256 -> 13 x13 x 5120.399 BFLOPs
15 conv 2551 x 1 / 1 13 x13 x 512 -> 13 x13 x 2550.044 BFLOPs
16 yolo
17 route13
18 conv 1281 x 1 / 1 13 x13 x 256 -> 13 x13 x 1280.011 BFLOPs
19 upsample 2x 13 x13 x 128 -> 26 x26 x 128
20 route19 8
21 conv 2563 x 3 / 1 26 x26 x 384 -> 26 x26 x 2561.196 BFLOPs
22 conv 2551 x 1 / 1 26 x26 x 256 -> 26 x26 x 2550.088 BFLOPs
23 yolo
Loading weights from yolov3-tiny.weights...Done!
data/dog.jpg: Predicted in 0.873869 seconds.
dog: 57%
car: 52%
truck: 56%
car: 62%
bicycle: 59%
zhaomeijing@raspberrypi5:~/workspace/04_yolo/darknet$ ./darknet detector test cfg/coco.data cfg/yolov3.cfg yolov3.weights data/dog.jpg
layer filters size input output
0 conv 323 x 3 / 1 608 x 608 x 3 -> 608 x 608 x320.639 BFLOPs
1 conv 643 x 3 / 2 608 x 608 x32 -> 304 x 304 x643.407 BFLOPs
2 conv 321 x 1 / 1 304 x 304 x64 -> 304 x 304 x320.379 BFLOPs
3 conv 643 x 3 / 1 304 x 304 x32 -> 304 x 304 x643.407 BFLOPs
4 res 1 304 x 304 x64 -> 304 x 304 x64
5 conv 1283 x 3 / 2 304 x 304 x64 -> 152 x 152 x 1283.407 BFLOPs
6 conv 641 x 1 / 1 152 x 152 x 128 -> 152 x 152 x640.379 BFLOPs
7 conv 1283 x 3 / 1 152 x 152 x64 -> 152 x 152 x 1283.407 BFLOPs
8 res 5 152 x 152 x 128 -> 152 x 152 x 128
9 conv 641 x 1 / 1 152 x 152 x 128 -> 152 x 152 x640.379 BFLOPs
10 conv 1283 x 3 / 1 152 x 152 x64 -> 152 x 152 x 1283.407 BFLOPs
11 res 8 152 x 152 x 128 -> 152 x 152 x 128
12 conv 2563 x 3 / 2 152 x 152 x 128 -> 76 x76 x 2563.407 BFLOPs
13 conv 1281 x 1 / 1 76 x76 x 256 -> 76 x76 x 1280.379 BFLOPs
14 conv 2563 x 3 / 1 76 x76 x 128 -> 76 x76 x 2563.407 BFLOPs
15 res 12 76 x76 x 256 -> 76 x76 x 256
16 conv 1281 x 1 / 1 76 x76 x 256 -> 76 x76 x 1280.379 BFLOPs
17 conv 2563 x 3 / 1 76 x76 x 128 -> 76 x76 x 2563.407 BFLOPs
18 res 15 76 x76 x 256 -> 76 x76 x 256
19 conv 1281 x 1 / 1 76 x76 x 256 -> 76 x76 x 1280.379 BFLOPs
20 conv 2563 x 3 / 1 76 x76 x 128 -> 76 x76 x 2563.407 BFLOPs
21 res 18 76 x76 x 256 -> 76 x76 x 256
22 conv 1281 x 1 / 1 76 x76 x 256 -> 76 x76 x 1280.379 BFLOPs
23 conv 2563 x 3 / 1 76 x76 x 128 -> 76 x76 x 2563.407 BFLOPs
24 res 21 76 x76 x 256 -> 76 x76 x 256
25 conv 1281 x 1 / 1 76 x76 x 256 -> 76 x76 x 1280.379 BFLOPs
26 conv 2563 x 3 / 1 76 x76 x 128 -> 76 x76 x 2563.407 BFLOPs
27 res 24 76 x76 x 256 -> 76 x76 x 256
28 conv 1281 x 1 / 1 76 x76 x 256 -> 76 x76 x 1280.379 BFLOPs
29 conv 2563 x 3 / 1 76 x76 x 128 -> 76 x76 x 2563.407 BFLOPs
30 res 27 76 x76 x 256 -> 76 x76 x 256
31 conv 1281 x 1 / 1 76 x76 x 256 -> 76 x76 x 1280.379 BFLOPs
32 conv 2563 x 3 / 1 76 x76 x 128 -> 76 x76 x 2563.407 BFLOPs
33 res 30 76 x76 x 256 -> 76 x76 x 256
34 conv 1281 x 1 / 1 76 x76 x 256 -> 76 x76 x 1280.379 BFLOPs
35 conv 2563 x 3 / 1 76 x76 x 128 -> 76 x76 x 2563.407 BFLOPs
36 res 33 76 x76 x 256 -> 76 x76 x 256
37 conv 5123 x 3 / 2 76 x76 x 256 -> 38 x38 x 5123.407 BFLOPs
38 conv 2561 x 1 / 1 38 x38 x 512 -> 38 x38 x 2560.379 BFLOPs
39 conv 5123 x 3 / 1 38 x38 x 256 -> 38 x38 x 5123.407 BFLOPs
40 res 37 38 x38 x 512 -> 38 x38 x 512
41 conv 2561 x 1 / 1 38 x38 x 512 -> 38 x38 x 2560.379 BFLOPs
42 conv 5123 x 3 / 1 38 x38 x 256 -> 38 x38 x 5123.407 BFLOPs
43 res 40 38 x38 x 512 -> 38 x38 x 512
44 conv 2561 x 1 / 1 38 x38 x 512 -> 38 x38 x 2560.379 BFLOPs
45 conv 5123 x 3 / 1 38 x38 x 256 -> 38 x38 x 5123.407 BFLOPs
46 res 43 38 x38 x 512 -> 38 x38 x 512
47 conv 2561 x 1 / 1 38 x38 x 512 -> 38 x38 x 2560.379 BFLOPs
48 conv 5123 x 3 / 1 38 x38 x 256 -> 38 x38 x 5123.407 BFLOPs
49 res 46 38 x38 x 512 -> 38 x38 x 512
50 conv 2561 x 1 / 1 38 x38 x 512 -> 38 x38 x 2560.379 BFLOPs
51 conv 5123 x 3 / 1 38 x38 x 256 -> 38 x38 x 5123.407 BFLOPs
52 res 49 38 x38 x 512 -> 38 x38 x 512
53 conv 2561 x 1 / 1 38 x38 x 512 -> 38 x38 x 2560.379 BFLOPs
54 conv 5123 x 3 / 1 38 x38 x 256 -> 38 x38 x 5123.407 BFLOPs
55 res 52 38 x38 x 512 -> 38 x38 x 512
56 conv 2561 x 1 / 1 38 x38 x 512 -> 38 x38 x 2560.379 BFLOPs
57 conv 5123 x 3 / 1 38 x38 x 256 -> 38 x38 x 5123.407 BFLOPs
58 res 55 38 x38 x 512 -> 38 x38 x 512
59 conv 2561 x 1 / 1 38 x38 x 512 -> 38 x38 x 2560.379 BFLOPs
60 conv 5123 x 3 / 1 38 x38 x 256 -> 38 x38 x 5123.407 BFLOPs
61 res 58 38 x38 x 512 -> 38 x38 x 512
62 conv 10243 x 3 / 2 38 x38 x 512 -> 19 x19 x10243.407 BFLOPs
63 conv 5121 x 1 / 1 19 x19 x1024 -> 19 x19 x 5120.379 BFLOPs
64 conv 10243 x 3 / 1 19 x19 x 512 -> 19 x19 x10243.407 BFLOPs
65 res 62 19 x19 x1024 -> 19 x19 x1024
66 conv 5121 x 1 / 1 19 x19 x1024 -> 19 x19 x 5120.379 BFLOPs
67 conv 10243 x 3 / 1 19 x19 x 512 -> 19 x19 x10243.407 BFLOPs
68 res 65 19 x19 x1024 -> 19 x19 x1024
69 conv 5121 x 1 / 1 19 x19 x1024 -> 19 x19 x 5120.379 BFLOPs
70 conv 10243 x 3 / 1 19 x19 x 512 -> 19 x19 x10243.407 BFLOPs
71 res 68 19 x19 x1024 -> 19 x19 x1024
72 conv 5121 x 1 / 1 19 x19 x1024 -> 19 x19 x 5120.379 BFLOPs
73 conv 10243 x 3 / 1 19 x19 x 512 -> 19 x19 x10243.407 BFLOPs
74 res 71 19 x19 x1024 -> 19 x19 x1024
75 conv 5121 x 1 / 1 19 x19 x1024 -> 19 x19 x 5120.379 BFLOPs
76 conv 10243 x 3 / 1 19 x19 x 512 -> 19 x19 x10243.407 BFLOPs
77 conv 5121 x 1 / 1 19 x19 x1024 -> 19 x19 x 5120.379 BFLOPs
78 conv 10243 x 3 / 1 19 x19 x 512 -> 19 x19 x10243.407 BFLOPs
79 conv 5121 x 1 / 1 19 x19 x1024 -> 19 x19 x 5120.379 BFLOPs
80 conv 10243 x 3 / 1 19 x19 x 512 -> 19 x19 x10243.407 BFLOPs
81 conv 2551 x 1 / 1 19 x19 x1024 -> 19 x19 x 2550.189 BFLOPs
82 yolo
83 route79
84 conv 2561 x 1 / 1 19 x19 x 512 -> 19 x19 x 2560.095 BFLOPs
85 upsample 2x 19 x19 x 256 -> 38 x38 x 256
86 route85 61
87 conv 2561 x 1 / 1 38 x38 x 768 -> 38 x38 x 2560.568 BFLOPs
88 conv 5123 x 3 / 1 38 x38 x 256 -> 38 x38 x 5123.407 BFLOPs
89 conv 2561 x 1 / 1 38 x38 x 512 -> 38 x38 x 2560.379 BFLOPs
90 conv 5123 x 3 / 1 38 x38 x 256 -> 38 x38 x 5123.407 BFLOPs
91 conv 2561 x 1 / 1 38 x38 x 512 -> 38 x38 x 2560.379 BFLOPs
92 conv 5123 x 3 / 1 38 x38 x 256 -> 38 x38 x 5123.407 BFLOPs
93 conv 2551 x 1 / 1 38 x38 x 512 -> 38 x38 x 2550.377 BFLOPs
94 yolo
95 route91
96 conv 1281 x 1 / 1 38 x38 x 256 -> 38 x38 x 1280.095 BFLOPs
97 upsample 2x 38 x38 x 128 -> 76 x76 x 128
98 route97 36
99 conv 1281 x 1 / 1 76 x76 x 384 -> 76 x76 x 1280.568 BFLOPs
100 conv 2563 x 3 / 1 76 x76 x 128 -> 76 x76 x 2563.407 BFLOPs
101 conv 1281 x 1 / 1 76 x76 x 256 -> 76 x76 x 1280.379 BFLOPs
102 conv 2563 x 3 / 1 76 x76 x 128 -> 76 x76 x 2563.407 BFLOPs
103 conv 1281 x 1 / 1 76 x76 x 256 -> 76 x76 x 1280.379 BFLOPs
104 conv 2563 x 3 / 1 76 x76 x 128 -> 76 x76 x 2563.407 BFLOPs
105 conv 2551 x 1 / 1 76 x76 x 256 -> 76 x76 x 2550.754 BFLOPs
106 yolo
Loading weights from yolov3.weights...Done!
data/dog.jpg: Predicted in 21.122667 seconds.
dog: 100%
truck: 92%
bicycle: 99%
zhaomeijing@raspberrypi5:~/workspace/04_yolo/darknet$2.2 测试结果展示及分析yolov3-tiny测试结果：yolov3测试结果：
序号名称精度时间（单位：秒）
1yolov3-tiny低快：0.873869
2.yolov3高慢：21.122667
通过测试结果可以得知yolov3的精度更高但是耗时，yolov3-tiny的速度最快但是精度不高。
//------end
页:
[1]