Advice on image resizing with minimal impact on FPS

I am a user developing an object detection and tracking system meant to be deployed on an edge on live video data. For this, resolution is likely at least 1080p in size, and the objection detection would need to run at a minimum speed (8 fps), my hardware is AGX Xavier, and I plan to deploy using TensorRT and via the Deepstream framework.

From Programming Comments - Darknet, OpenCV, and FPS, running darknet+cuda on AGX Xavier can lead to a 4.8x difference depending on whether the images are preprocessed or need to be resized. As we are planning to run Yolov3/Yolov4/Scaled-Yolov4, the models would likely be quite large and we would require some downsizing of the images.

I was wondering what are the ways to implement this image resizing in the fastest way on a video stream? We expect a standardized image stream resolution (all 1080p, or some standard higher resolution on any single camera). It will also only be one video stream. The frames will be passed to an object detector and then object tracker. I’ve looked around and seen some posts such as Resize image array with GStreamer+opencv in Jetson nano using hardware scaling - #3 by DaneLLL, but I’m not sure.

We have demonstration of running YoloV3, YoloV4 in DeepStream SDK. Please take a look at

/opt/nvidia/deepstream/deepstream-6.0/sources/objectDetector_Yolo$ ls
config_infer_primary_yoloV2_tiny.txt   deepstream_app_config_yoloV3.txt
config_infer_primary_yoloV2.txt        labels.txt
config_infer_primary_yoloV3_tiny.txt   nvdsinfer_custom_impl_Yolo
deepstream_app_config_yoloV2_tiny.txt  README
deepstream_app_config_yoloV2.txt       yolov3-calibration.table.trt7.0

GitHub - NVIDIA-AI-IOT/yolov4_deepstream

Please try the demonstrations. For resizing image, you can use nvvideoconvert plugin. Please check
Gst-nvvideoconvert — DeepStream 6.0 Release documentation

All documents are in
NVIDIA Metropolis Documentation

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.