Detection Issues with Cylindrical Object in Custom Dataset

[Septomide]

Hi,

I am working with a custom dataset containing multiple objects. Most objects are detected well, but I am experiencing issues with the detection of a cylindrical object. It seems to perform poorly, likely due to symmetry-related challenges.

roi, roi_mask_trunc, roi_region, .. look the same like they do for other objects. The sym_info is set for this object, and is of shape (942, 3, 3). The sym_info set in the models_info.json:

{
  # rotation around z, cylinder is upright
  "symmetries_continuous": [
    {
      "axis": [0, 0, 1],
      "offset": [0, 0, 0]
    }
  ],
  # rotation 180 around x and y 
  "symmetries_discrete": [ 
    [
      1, 0, 0, 0,
      0, -1, 0, 0,
      0, 0, -1, 0,
      0, 0, 0, 1
    ],
    [
      -1, 0, 0, 0,
      0, 1, 0, 0,
      0, 0, -1, 0,
      0, 0, 0, 1
    ]
  ]
}

Image samples given from the dataloader of the cylinder object:

Results after training (center, rotation, and scale of the cylinder are completely off):

Did I miss anything obvious? Any guidance on improving detection for symmetrical objects would be greatly appreciated.

[shanice-l]

You can try to verify the sym_info set using visualization scripts.
e.g. You can transform the point cloud with sym_info, and to see if it is matched with original point cloud.