FPS Drop due to OCR Detection RTX

Gaming and Visualization Technologies General Topics and Other SDKs
, , ,
1

We are experiencing a significant FPS drop when using OCR for detecting vehicle number plates and other objects. We have used paddle OCR for detecting the number plates.

The specifications of the GPU we are using are listed below:

OMEN by HP 45L Gaming Desktop GT22-1214
13th Generation Intel® Core™ i9 processor
Windows 11 Pro
Kingston FURY 64 GB DDR5-4400 MHz XMP RGB Heatsink RAM
1 TB WD Black PCIe® NVMe™ TLC M.2 SSD
2 TB 7200 rpm SATA HDD
NVIDIA® GeForce RTX™ 4090

We also attempted multi-processing

2

Could you share more information including:

  • Logs from NVIDIA SMI using commands: nvidia-smi dmon
  • Workloads running on GPU concurrently including projects and ways to reproduce.

Additionally, we don’t recommend running multiple applications(Multi-process) on one GPU, better running with multiple threads. We have DeepStream SDK which provides license platter recognition demo here: GitHub - NVIDIA-AI-IOT/deepstream_lpr_app: Sample app code for LPR deployment on DeepStream

3

OCR_STATS
OCR_STATS866×700 26 KB

Total Number Cameras - 17
Avg FPS per camera - 4 to 7
Inference details - Pytorch ONNX

4

According to the log, the GPU isn’t fully utilized. I’d suggest starting with Deepstream SDK for video analytics tasks.