I premise that I compiled the package by using "colcon build --packages-select isaac_ros_dope".
I'm facing some problems with this package while running on a NVIDIA Jetson Orin Nano.
More in detail, I think there could be some issues with the intrinsics of the camera.
I'm using the package with the images provided by an Intel RealSense d435i that is configured with "rgb_camera.profile:=640x480x30".
I'm quite sure that the node "dope_decoder_node" is loading the camera calibration parameters that I specified in the "dope_config.yaml" file.
The calibration parameters specified in the "dope_config.yaml" are the following: [605.6713256835938, 0, 329.4186706542969, 0, 605.2176513671875, 245.733154296875, 0, 0, 1]. These parameters are the same of those in the "camera_info" topic.
The package is running quite well in the Jetson Orin Nano but the estimated position of the object that I'm detecting (Campbell's Tomato soup) looks too close to the camera.
The object is approximately 20 cm far from the camera but DOPE is estimating a distance of something like 10 cm.
Moreover, I recorded a bag file with the following topics: images, camera_info, and poses. In my laptop, I converted the bag file to be used in ROS1 Noetic. Subsequently, I used the bag file to run DOPE (https://github.com/NVlabs/Deep_Object_Pose) and compare the results.
As you can see from the following image, the results provided by the isaac_ros_dope and the ROS1 DOPE are different.
Thank you in advance for your help.
PS. In the Jetson Orin Nano, I'm using isaac_ros_dope with the weights converted in the ONNX format and with the "isaac_ros_dope_tensor_rt.launch.py" file.
I premise that I compiled the package by using "colcon build --packages-select isaac_ros_dope". I'm facing some problems with this package while running on a NVIDIA Jetson Orin Nano. More in detail, I think there could be some issues with the intrinsics of the camera. I'm using the package with the images provided by an Intel RealSense d435i that is configured with "rgb_camera.profile:=640x480x30". I'm quite sure that the node "dope_decoder_node" is loading the camera calibration parameters that I specified in the "dope_config.yaml" file. The calibration parameters specified in the "dope_config.yaml" are the following: [605.6713256835938, 0, 329.4186706542969, 0, 605.2176513671875, 245.733154296875, 0, 0, 1]. These parameters are the same of those in the "camera_info" topic. The package is running quite well in the Jetson Orin Nano but the estimated position of the object that I'm detecting (Campbell's Tomato soup) looks too close to the camera. The object is approximately 20 cm far from the camera but DOPE is estimating a distance of something like 10 cm. Moreover, I recorded a bag file with the following topics: images, camera_info, and poses. In my laptop, I converted the bag file to be used in ROS1 Noetic. Subsequently, I used the bag file to run DOPE (https://github.com/NVlabs/Deep_Object_Pose) and compare the results. As you can see from the following image, the results provided by the isaac_ros_dope and the ROS1 DOPE are different.
Thank you in advance for your help.
PS. In the Jetson Orin Nano, I'm using isaac_ros_dope with the weights converted in the ONNX format and with the "isaac_ros_dope_tensor_rt.launch.py" file.