Yolo sample with non-squared network input

Accelerated Computing Intelligent Video Analytics DeepStream SDK
#1

Hi all,

Is it possible to run yolo object detection demo with non-squared input (not 320x320/416x416/608x608 but 320x608 for example)? To run it “non-squared” with pure darknet yolo implementation you can only change yolov3.cfg parameters as shown below:

[net]
width=608
height=320

But in deepstream-app sample there are a lot of asserts appear if width!=height.
Is it just sample limitation or tensorRT limitation?

#2

Hi,

This is only a sample limitation. The sample currently supports just the standard yolo models - yolov2, yolov2-tiny, yolov3 and yolov3-tiny. You can modify the network builder part of the code to suit your custom model architecture.

#3