[BUG] Detection output on mono camera looks not to scale

[MRo47]

Version: 1 commit behind iron commit hash cf5d2aaee9117298ea1632c98ffd36a5d7d535ac

Issue

• Trying to run a Depth pipeline on OAK-D Pro W camera.
• What I want is depth output with a yolo detection running on the left camera.
• The detections look they are not to scale, is there a parameter I need to set to scale them?

Steps to reproduce

1. Set config (camera.yaml) as below

   /**:
   ros__parameters:
   camera:
     i_enable_imu: true
     i_enable_ir: true
     i_nn_type: none
     i_pipeline_type: Depth
   left:
     i_publish_topic: true
     i_enable_nn: true
     i_disable_node: false
     i_resolution: '720P'
   left_nn:
     i_board_socket_id: 1
     i_nn_config_path: depthai_ros_driver/yolo

2. Modify launch file to add the depthai_filters::Detection2DOverlay and detection_labels, these have to be set up for the overlay as well as the default are mobilenet ssd labels.

detection_labels = [
        "person",
        "bicycle",
        "car",
        "motorbike",
        "aeroplane",
        "bus",
        "train",
        "truck",
        "boat",
        "traffic light",
        "fire hydrant",
        "stop sign",
        "parking meter",
        "bench",
        "bird",
        "cat",
        "dog",
        "horse",
        "sheep",
        "cow",
        "elephant",
        "bear",
        "zebra",
        "giraffe",
        "backpack",
        "umbrella",
        "handbag",
        "tie",
        "suitcase",
        "frisbee",
        "skis",
        "snowboard",
        "sports ball",
        "kite",
        "baseball bat",
        "baseball glove",
        "skateboard",
        "surfboard",
        "tennis racket",
        "bottle",
        "wine glass",
        "cup",
        "fork",
        "knife",
        "spoon",
        "bowl",
        "banana",
        "apple",
        "sandwich",
        "orange",
        "broccoli",
        "carrot",
        "hot dog",
        "pizza",
        "donut",
        "cake",
        "chair",
        "sofa",
        "pottedplant",
        "bed",
        "diningtable",
        "toilet",
        "tvmonitor",
        "laptop",
        "mouse",
        "remote",
        "keyboard",
        "cell phone",
        "microwave",
        "oven",
        "toaster",
        "sink",
        "refrigerator",
        "book",
        "clock",
        "vase",
        "scissors",
        "teddy bear",
        "hair drier",
        "toothbrush",
    ]

    detection_viz_node = ComposableNode(
        package="depthai_filters",
        plugin="depthai_filters::Detection2DOverlay",
        parameters=[
            {"label_map": detection_labels},
        ],
        remappings=[
            ("rgb/preview/image_raw", "/oak/left/image_raw"),
            ("nn/detections", "/oak/left_nn/detections"),
        ],
    )

    ...

    ComposableNodeContainer(
            name=f"{name}_container",
            namespace=namespace,
            package="rclcpp_components",
            executable="component_container",
            composable_node_descriptions=[
                ComposableNode(
                    package="depthai_ros_driver",
                    plugin="depthai_ros_driver::Camera",
                    name=name,
                    namespace=namespace,
                    parameters=[
                        params_file,
                        tf_params,
                        parameter_overrides,
                        {"left_nn.i_label_map": detection_labels}
                    ],
                ),
                detection_viz_node,
            ],
            arguments=["--ros-args", "--log-level", log_level],
            prefix=[launch_prefix],
            output="both",
        ),

3. Build and launch

   ros2 launch depthai_ros_driver camera.launch.py

Below is how the overlay looks, the person doesn't look like this on camera, trust me and we don't have ghosts ;)

[MRo47]

scaling_issue.zip

Complete launch file and yaml config

[Serafadam]

Hi, thanks for the report, indeed there is a minor bug here, fix will be on the way but just to recap:

• Images fed to the NN node need to be resized to fit the NN input, in this case 416x416, by default this is automatically done by an internal image manip node, the input image is "squeezed" to preserve FOV of the original image, more information on that can be found here
• Since the image is squeezed, the resulting bounding boxes are also "squeezed" to fit this format
• You can verify that by using the passthrough output from NN node, can be enabled with left_nn.i_enable_passthrough: true, when you pass that image to the detection overlay you should see the bounding boxes aligned to the squeezed image
• DetectionOverlay2D filter is missing the desqueeze parameter which would apply correct transformation, this will be fixed to mimic the behavior in spatial_bb, this is a small modification so if you have built depthai-ros from source you can make the modification or wait for the next release.

[MRo47]

@Serafadam Thank you for the pointers above.

Yup that was my guess that it was the scaling done my manip node before input to the NN and I can confirm that on setting left_nn.i_enable_passthrough: true and doing the overlay on the output topic the output is correctly aligned.

From a design perspective I have 3 questions

1. Shouldn't desqueeze be performed on the Depth node directly? My reasoning is that the user gave the model input of size WxH so they should expect the detections to be of same size right?
2. Would the above complicate the implementation somehow?
3. To perform desqueeze on DetectionOverlay2D filter node you will need to have information on the "squeeze ratios" meaning input_image_size:NN_input_size, are these captured somewhere in a topic? or would this require passing the NN configs to the DetectionOverlay2D filter?

[Serafadam]

An implementation of this dequeeze is available on this branch to test out. A small modification to the DetectionOverlayNode in launch file is needed for it:

remappings=[ 
("rgb/preview/image_raw", "/oak/left/image_raw"),
("rgb/preview/camera_info", "/oak/left_nn/camera_info"), 
("nn/detections", "/oak/left_nn/detections"), 
 ], 

Shouldn't desqueeze be performed on the Depth node directly? My reasoning is that the user gave the model input of size WxH so they should expect the detections to be of same size right?

At this moment this is not implemented in the API, this is the default behavior in both C++ and Python API as described in the documentation. This might change in the future.

To perform desqueeze on DetectionOverlay2D filter node you will need to have information on the "squeeze ratios" meaning input_image_size:NN_input_size, are these captured somewhere in a topic? or would this require passing the NN configs to the DetectionOverlay2D filter?

Currently it is done based on the camera_info topic taken from passthrough image, unfortunately ROS Vision message doesn't carry this information

[MRo47]

Hey @Serafadam

I was able to have a workaround for this in #606