search

PINTO0309 / TPU-Posenet

Edge TPU Accelerator / Multi-TPU / Multi-Model + Posenet/DeeplabV3/MobileNet-SSD + Python + Sync / Async + LaptopPC / RaspberryPi
https://qiita.com/PINTO
MIT License
42 stars 14 forks source link

about FPS on raspberry pi #2

Closed willfu closed 4 years ago

willfu commented 4 years ago

@PINTO0309 thanks for the work. I am curious about the frame rate on raspberry pi. do you have any data for that? the demo video are all from PC, right?

PINTO0309 commented 4 years ago

Yes. Only the results verified on the PC were posted in the README. I got Pi4 the other day so I will try to verify it when I have time.

willfu commented 4 years ago

thanks for the reply, is CPU the bottleneck? or the io? as you are the expert ;)

PINTO0309 commented 4 years ago

To put it simply, it is USB2.0 communication speed and CPU performance. Although it is a Japanese article, I would like to introduce the verification article of my acquaintance that I think is useful for you.

Edge TPU USB Accelerator analysis-I/O data transfer - Qiita - iwatake2222

Analysis of Edge TPU USB Accelerator-Operation and model structure - Qiita - iwatake2222

willfu commented 4 years ago

Thanks for the introduction, which is based on jetson nano and tpu, I will also have a test.

PINTO0309 commented 4 years ago

Pi4 + TPU x3 (Deeplabv3 + PoseNet + MobileNetV2-SSD) The PiCamera settings don't seem to work. The shooting speed has become very slow... ezgif com-video-to-gif

willfu commented 4 years ago

thanks for the sharing. is this the result of tpux3 ? very strange...

PINTO0309 commented 4 years ago

This is a sample of TPUx3. Because the camera shooting rate is only 10 FPS, the camera shooting speed is slower than inference. I think it works faster if the camera is set up correctly. Note that when 3 TPUs are used, all 4 Cores of the CPU are used up, so the CPU pre-processing and post-processing overhead is the bottleneck.

willfu commented 4 years ago

ok. Got it. thanks.