Run PoseNet on RaspberryPi 4 and compare edge device performance (Bonus: Jetson Xavier)
【apology】
The contents have been updated to those remeasured by changing the power supply. According to the record initially released, it seems that the power supply of the Raspberry Pi 4 was defective, it was not able to supply enough power, and the performance was degraded. I confirmed that the clock frequency is rotating at the maximum frequency, but ...
[2019/12/03 postscript]
I touched Jetson AGX Xavier, so I added the data. The method of installing EgdeTPU in Jetson Xavier was the same as Jetson Nano. Please refer to the following article for details.
[Using EdgeTPU with Jetson Nano] [Jetson Nano + EdgeTPU with explosive PoseNet (performance comparison with Raspberry Pi)]
[Details]
** This article is a result only. ** **
Just as I measured the performance of PoseNet in in the past article, I also measured the performance of Raspberry Pi 4.
【result】
[Execution result (camera image)]
| resolution | Platform | Inference time(ms) | FrameIO (ms) | FPS | Remarks |
|---|---|---|---|---|---|
| 1280x720 | RaspberryPi3 + EdgeTPU | 279.4 | 75.82 | 2.47 | |
| RaspberryPi4 + EdgeTPU | 44.9 | 54.14 | 6.95 | ||
| - | DevBoard | 44.9 | 42.89 | 7.46 | |
| - | JetsonNano + EdgeTPU | 49.2 | 23.82 | 7.44 | |
| - | JetsonXavier + EdgeTPU | 42.5 | 6.91 | 7.40 | |
| 640x480 | RaspberryPi3 + EdgeTPU | 94.4 | 27.48 | 7.16 | |
| RaspberryPi4 + EdgeTPU | 15.0 | 18.96 | 19.54 | ||
| - | DevBoard | 13.5 | 15.26 | 24.36 | |
| - | JetsonNano + EdgeTPU | 15.0 | 8.53 | 29.89 | Camera performance limits |
| - | JetsonXavier + EdgeTPU | 13.7 | 5.62 | 29.98 | Camera performance limits |
| 480x360 | RaspberryPi3 + EdgeTPU | 53.5 | 12.34 | 12.57 | |
| - | RaspberryPi4 + EdgeTPU | 8.0 | 4.24 | 29.98 | Camera performance limits |
| - | DevBoard | 8.2 | 10.92 | 30.18 | Camera performance limits |
| - | JetsonNano + EdgeTPU | 9.7 | 4.21 | 30.06 | Camera performance limits |
| - | JetsonXavier + EdgeTPU | 8.2 | 2.18 | 29.98 | Camera performance limits |
[Result (video file)]
| resolution | Platform | Inference time(ms) | FrameIO (ms) | FPS | Remarks |
|---|---|---|---|---|---|
| 640x480 | RaspberryPi3 + EdgeTPU | 65.3 | 19.59 | 8.68 | |
| - | RaspberryPi4 + EdgeTPU | 14.8 | 9.53 | 23.03 | |
| - | DevBoard | 13.6 | 14.97 | 20.94 | |
| - | JetsonNano + EdgeTPU | 15.0 | 6.87 | 30.08 | |
| - | JetsonXavier + EdgeTPU | 13.3 | 2.33 | 43.70 | |
| 480x360 | RaspberryPi3 + EdgeTPU | 33.9 | 10.27 | 14.64 | |
| - | RaspberryPi4 + EdgeTPU | 8.1 | 4.17 | 38.44 | |
| - | DevBoard | 7.5 | 8.18 | 31.68 | |
| - | JetsonNano + EdgeTPU | 8.9 | 4.00 | 45.57 | |
| - | JetsonXavier + EdgeTPU | 7.4 | 1.51 | 67.04 |
Initially, the performance was not as good as I expected due to insufficient power supply capacity, which was a disappointing result, but when I changed to a power supply that can supply 3A firmly, the score was comparable to DevBoard. Some processes exceed DevBoard. Comparing the Raspberry Pi 4 and DevBoard, which have such performance and 4GB of memory, I feel that the Raspberry Pi 4 is superior in terms of cost performance.
Will DevBoard end up in romance ... DevBoard seems to have been updated with various functions the other day, so I will remeasure it with the latest version at a later date.