I’m trying to get suitable software for face-detection using a VIM3 device. So, I did two experiment cases.
The first is an example provided by KHADAS. (https://github.com/khadas/ksnn, I ran yolov3-face.py)
This demo resulted in about 40% CPU usage and about 8~10 FPS.
Restricted the figure attachment to new user issues
Second, it is open source. (https://github.com/Saadmairaj/face-detectors, I used UL320)
This demo resulted in about 300% CPU usage and about 15~17 FPS.
By the way, I don’t know if I used the NPU enough for the second case.
I would like to implement the intermediate performance between the two results. In other words, I’m expecting an increase in terms of FPS with more CPU usage than the first result. Is there such a way? Please help. @numbqq@Frank
@mwjung Is this the result of you running it directly on the cpu? If it is, then when you start to use NPU, the frame rate will rise. The CPU usage mainly depends on how your application handles it, and it may not be higher than ksnn’s demo.
@Frank I contacted the source code author of the second experiment, and he said that it does not support the use of NPU. But he said that onnxruntime is used.
By the way, I want to confirm whether KSNN’s demo can expect a rise in FPS by using more system resources.
I think you don’t understand how the NPU is used, all the original models do not support the use of the NPU, and the original model must be converted into a format that can be used on the NPU through a conversion tool. NPU can only be used, so all models in KSNN are converted
You can use the conversion tool to convert and then call it through the NPU. If you just run on the CPU, there is no way to provide more system resources to increase the frame rate
