Custom object detection non-square aspect ratio

Software ROS2
1

Hello there,

we have successfully deployed our custom Yolo model using the zed_ros2_wrapper + isaac_ros_image_pipeline + isaac_ros_dnn_inference. The model is trained on 1280x800 with the isaac_ros_image_pipeline scaling the 1920x1200 native resolution of the ZEDX down to the model input size.

We now want to try running the model directly from the zed_ros2_wrapper and at 1920x1280, but as far as I can tell custom object detection is only possible with a square input size? Also, is it possible to define a ROI to be used for object detection e.g. 1920x800?

2

I looked at the way the custom detector onnx_yolo_internal example deals with the input resolution an realized that the zed_ros2_wrapper is shadowing the capability to set sl::ObjectDetectionParameters.custom_onnx_dynamic_input_shape to a non-square sl::Resolution. I created a feature request to address this on github.