Object Detection with yolo v1, opencv-3.4.6, python

Hi,

Currently, I am working on jetson nano, and I am facing problem while using RSTP(live streaming)@20FPS using OpenCV.

1)Does OpenCV utilizes GPU of jetson nano?
2)Processing results are too slow for live streaming even though GPU utilization is 99% and CPU usage is below 30-40% (4 Cores). Why this happens so?
3)Which prebuilt version of OpenCV jetson nano supports if it jetson nano has inbuilt OpenCV library why it doesn’t support gstream which I get to know from(https://jkjung-avt.github.io/opencv-on-nano/).
4)What is the benchmark for tiny-Yolo v1?

Hi,

1. Depends on the API you used.
Are you reading the RTSP source with GStreamer pipeline?
If yes, you can check this document to create a hardware accelerated pipeline:
https://developer.download.nvidia.com/embedded/L4T/r32_Release_v1.0/Docs/Accelerated_GStreamer_User_Guide.pdf

You can decide to execute each component on CPU or GPU based on the module/parameter you choose.

2. This depends on the API you used.
In general, there are lots of memory copy in OpenCV, which make the whole pipeline slow.
To solve this issue, we provide an optimized library for multimedia deep learning jobs:
https://developer.nvidia.com/deepstream-sdk

3. The prebuilt OpenCV is v3.3.1.
GStreamer is a third-party library and there are some legal issue to enable it by default.

4. We don’t benchmark YOLOv1 but here are some score for YOLOv3 for your reference:
https://developer.nvidia.com/embedded/jetson-nano-dl-inference-benchmarks

Thanks.

Why any benchmarks are not given based on python?
Is there any common source which measures performance on board with DeepStream and without DeepStream?