We have been performing OCR detection for vehicle number plates and beams processed in galvanization plants. Our model has been benchmarked on a top-tier device, such as the Orin NX, where we can run 8 cameras at 200 FPS, achieving 20 to 25 FPS per camera. However, when we perform OCR recognition using the PaddleOCR module, the FPS gradually decreases to 8 FPS. Moreover, this 8 FPS is achievable only if there is one beam; with multiple beams or vehicles, the FPS drops to 2 to 4. We are seeking expert suggestions or guidance to solve this problem. This issue is not limited to edge devices; we face the same performance drop with high-end GPU systems like the NVIDIA RTX 40 series
Hi,
Could you share the tegrastats output with us?
This can help us know more about the utilization in your use case.
$ sudo tegrastats
Thanks.
can i share you the jetson-stats?
Hi,
Yes, the log should be very similar.
Thanks.
Hi,
Your CPU usage is very high.
Could you try to enable the remaining 4 CPUs to see if this helps?
$ sudo nvpmodel -m 0
$ sudo jetson_clocks
Thanks.
Does Enabling the other CPU will result in higher FPS, Let me try and get back to you, Thanks for the reply
Can you explain about jetson clock and NVP model, what it will do
Hi,
nvpmodel is our provided power mode depending on the different power budgets.
The document shared above has listed the maximum clock of each mode (ex. CPU, GPU, VIC, …).
For jetson_clocks, it will fix the clock to the maximum to avoid the performance drop from dynamic frequency mechanism.
Thanks.
1 Like
This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.