OAK相机跑各种yolo模型的检测帧率和深度帧率 – OAKChina

OAK相机跑各种yolo模型的检测帧率和深度帧率

by OAK中国
in 其他
on 2023年2月17日

0

鉴于我们社群中有不少朋友在将YOLO模型部署到OAK相机上，大家对帧率比较关心，所以我们做了一个比较全的测试，仅供大家参考。

选用的模型有yolov5r7、yolov6r3、yolov7、yolov8，分别在320×320、416×416、640×640下的帧率。

▌测试环境

设备：OAK-D-S2

DepthAI：2.20.2.0

系统：Win10

测试的YOLO模型种类：

- yolov5r7 : https://github.com/ultralytics/yolov5/releases/tag/v7.0

- yolov6r3 : https://github.com/meituan/YOLOv6/releases/tag/0.3.0

- yolov7 : https://github.com/WongKinYiu/yolov7/releases/tag/v0.1

- yolov8 : https://github.com/ultralytics/assets/releases/tag/v0.0.0

结果说明：“/”表示没有此版本，“×”表示运行出问题（没有识别结果或运行一会中断），没填数据的表示没测试。

▌YOLO检测帧率结果

320x320	yolov5r7		yolov6r3	yolov7	yolov8
320x320	yolov3	yolov5	yolov6r3	yolov7	yolov3u	yolov5u	yolov8
Nano / Tiny	37.9 fps	42.5 fps	52.8 fps	37.7 fps	30.2 fps	36.2 fps	35.0 fps
Small	/	22.1 fps	24.1 fps	/	/	21.4 fps	17.7 fps
Medium	3.4 fps	10.2 fps	9.9 fps	/	2.4 fps	9.1 fps	7.2 fps
Large / SPP	3.5 fps	3.5 fps	3.4 fps	3.7 fps	2.4 fps	4.1 fps	3.5 fps
X	/	1.9 fps	/	2.0 fps	/	1.9 fps	1.9 fps

416x416	yolov5r7		yolov6r3	yolov7	yolov8
416x416	yolov3	yolov5	yolov6r3	yolov7	yolov3u	yolov5u	yolov8
Nano / Tiny	20.1 fps	29.5 fps	35.6 fps	23.1 fps	18.1 fps	23.8 fps	22.4 fps
Small	/	14.3 fps	13.1 fps	/	/	13.0 fps	10.8 fps
Medium	1.8 fps	5.5 fps	4.9 fps	/	1.2 fps	5.2 fps	4.2 fps
Large / SPP	1.8fps	2.2 fps	2.2 fps	1.9 fps	1.2 fps	2.1 fps	1.9 fps
X	/	1.3 fps	/	1.3 fps	无	1.2 fps	1.4 fps

640x640	yolov5r7		yolov6r3	yolov7	yolov8
640x640	yolov3	yolov5	yolov6r3	yolov7	yolov3u	yolov5u	yolov8
Nano / Tiny	8.0 fps	16.2 fps	18.6 fps	10.1 fps	6.3 fps	13.3 fps	12.3 fps
Small	/	5.1 fps	4.5 fps	/	/	5.8 fps	4.4 fps
Medium	0.8 fps	2.2 fps	2.1 fps	/	0.4 fps	2.0 fps	2.1 fps
Large / SPP	0.8 fps	1.3 fps	1.1 fps	1.1 fps	0.4 fps	1.1 fps	1.2 fps
X	/	0.6 fps	/	0.7 fps	/	0.6 fps	0.6 fps

▌YOLO检测+深度帧率结果

320x320	yolov5r7		yolov6r3	yolov7	yolov8
320x320	yolov3	yolov5	yolov6r3	yolov7	yolov3u	yolov5u	yolov8
Nano / Tiny	30.9 fps	35.4 fps	40.6 fps	27.1 fps	27.3 fps	30.6 fps	29.9 fps
Small	/	18.1 fps	21.0 fps	/	/	15.8 fps	11.4 fps
Medium	❌	❌	❌	/	0.4 fps	❌	❌
Large	❌	❌		❌	❌	❌	❌
X	/	❌	/	❌	/	❌	❌

416x416	yolov5r7		yolov6r3	yolov7	yolov8
416x416	yolov3	yolov5	yolov6r3	yolov7	yolov3u	yolov5u	yolov8
Nano / Tiny	17.5 fps	25.2 fps	27.2 fps	18.1 fps	10.7 fps	18.4 fps	16.3 fps
Small	/	❌	❌	/	/	❌	❌
Medium	❌	❌	❌	/	❌	❌	❌
Large	❌	❌	❌	❌	❌	❌	❌
X	/	❌	/	❌	/	❌	❌

640x640	yolov5r7		yolov6r3	yolov7	yolov8
640x640	yolov3	yolov5	yolov6r3	yolov7	yolov3u	yolov5u	yolov8
Nano / Tiny	❌	7.5 fps	10.5 fps	3.2 fps	❌	❌	❌
Small	/	❌	❌	/	/	❌	❌
Medium	❌	❌	❌	/	❌	❌	❌
Large	❌	❌	❌	❌	❌	❌	❌
X	/	❌	/	❌	/	❌	❌

▌结论

不管是做纯检测，还是检测+深度，最小的Nano/Tiny版本，跑出来帧率是最高的。分辨率越大，帧率越小。加了深度后，帧率也会降低。

Tags: #YOLO

×

目录