Parsing CenterPoint Output ONNX Nodes

[OrcunCanDeniz]

I managed to get the ONNX export for mmdet3d's centerpoint. However I am not sure how do I parse the outputs. For example, I have the ONNX model as in the screenshot, for a network with 3 tasks each with 1 class. Direction, 3D boxes and confidences from all tasks are concated at the output. How do i be sure that any given index or slice of these output tensors is associated to a specific task t and class c. What is the recommended approach to getting this kind of mapping info for each architecture, is it simply the same order in the cfg file maybe?

  pts_bbox_head=dict(
    type='CenterHead',
    in_channels=sum([128, 128, 128]),
    tasks=[
        dict(num_class=1, class_names=['car']),
        dict(num_class=1, class_names=['bicycle']),
        dict(num_class=1, class_names=['pedestrian']),
    ],
    common_heads=dict(
        reg=(2, 2), height=(1, 2), dim=(3, 2), rot=(2, 2)),

     ...

[tpoisonooo]

1. The output (dir_score, bbox_preds and scores) needs postprocess here: https://github.com/open-mmlab/mmdeploy/blob/1c5dcb799badbad717f6bfed615acb4a8e52f92f/mmdeploy/codebase/mmdet3d/models/centerpoint.py#L68

centerpoint__get_bbox is actually called here

Finally get_bbox() would return real bbox/score/label. The label value [0,1,2] maps to [car, bicyle, ped]

2. Furthermore, End2End centerpoint model (without postprocess) is still WIP.

[OrcunCanDeniz]

Thanks !