I am working with Isaac ROS pose estimation, but the results I get are less than desirable. Using the centerpose_shoe model, objects are detected less than 10% of the time without moving them. Using the ketchup DOPE model, I detect objects even less often.
My test objects are a bottle of French mustard, 3 Campbells’ tomato soup cans, a box of Cheez It and a can of Spam (basically the objects used to train DOPE).
I was therefore wondering if there were preferable conditions to run these models, like having a white background, lots of light, camera resolution, etc.
The DOPE models were trained from the HOPE dataset and should have been usable even if there were slight variations in the objects too. We’re checking with the NV Research team that developed DOPE if they have any guidance here. Could you post a screen shot of what your ketchup bottle scene looks like? What camera are you using and any other nodes of note in your ROS 2 graph?
I am using an Intel RealSense d435 3D camera to capture the images. Here is a picture of my testing area. For context, there is 60 cm between the camera and the white sheet.
Regarding my node graph, the only nodes in my graph are the realsense2_camera_node as well as the nodes launched with the command ros2 launch isaac_ros_dope isaac_ros_dope_triton.launch.py model_name:=Ketchup model_repository_paths:=['/src/models'] input_binding_names:=['INPUT__0'] output_binding_names:=['OUTPUT__0'] object_name:=Ketchup
I also wanted to inform you that I opened an issue on the issac ROS pose estimation GitHub repository so people don’t work twice.
