Run YOLO+TensorRT without deepstream

Hello,

If I want to know the performance effect of TensorRT, how to remove deepstream from deepstream-yolo-app?
and the remained YOLO+TensorRT can input video file and execute object detection.

I run deepstream-yolo-app Tegra sample_720p.h264 config/yolov3.txt work well,
my environments are Jetpack4.1.1, DeepStream3.0 and deepstream_reference_apps on Xavier.

Hi,

There are two app in the YOLO sample.
trt-yolo-app is the standalone sample without using deepstreamSDK.
https://github.com/NVIDIA-AI-IOT/deepstream_reference_apps#trt-yolo-app

Thanks.

@AastaLLL

I have run trt-yolo-app, but it can only input image files.
I need a standalone sample without using deepstreamSDK and it can be input video file.
Is there any sample program? or how to modify deepstream-yolo-app?

Hi,

You can update the source function here to support video input:
https://github.com/NVIDIA-AI-IOT/deepstream_reference_apps/blob/master/sources/apps/trt-yolo/trt-yolo-app.cpp#L124

Thanks.

Hi,could you tell me how to modify deepstream-yolo-app to use video file?

Hi,

Deepstream-yolo-app read input data with GStreamer.
You can update the GStreamer command to V4L2 source directly:
https://github.com/NVIDIA-AI-IOT/deepstream_reference_apps/blob/master/yolo/apps/deepstream-yolo/deepstream-yolo-app.cpp#L166

Thanks.

I’m sorry.I want to ask trt-yolo-app,not deepstream-yolo-app.

Hi,

You can try to read camera with OpenCV and update data as cvmat here:
https://github.com/NVIDIA-AI-IOT/deepstream_reference_apps/blob/master/yolo/apps/trt-yolo/trt-yolo-app.cpp#L115

Thanks.