Question about SAR-Net Results

[taeyeopl]

I am interested in your work, and I am trying to compare the results of your method (SAR-Net, CVPR22). Can you share your method .pkl files of Table 1 on the REAL275??

If you can share it quickly, it would be of great help. My email is taeyeop.trevor@gmail.com

[hetolin]

Thanks for your interested in our work. I have uploaded the .pkl files of SAR-Net on REAL dataset to the Google Drive.

[taeyeopl]

Thanks for your quick reply with the results.

I have checked the results, and it looks like your result['pred_RTs'] parameter doesn't consider scale, always, scale is one. I used a similar this code https://github.com/hughw19/NOCS_CVPR2019/blob/78a31c2026a954add1a2711286ff45ce1603b8ab/utils.py#L467

[hetolin]

Yes, the result['pred_RTs'] parameter doesn't consider scale, the predicted scale is stored in result['pred_scales']

[taeyeopl]

I initially considered the scale and looks fine. This is my visualization code, and please feel free to let me know if there are any mistakes

# Reference code 
 (https://github.com/hughw19/NOCS_CVPR2019/blob/78a31c2026a954add1a2711286ff45ce1603b8ab/utils.py#L2833)
#xyz_axis = 0.3*np.array([[0, 0, 0], [0, 0, 1], [0, 1, 0], [1, 0, 0]]).transpose()
xyz_axis = np.array([[0, 0, 0], [0, 0, pred_size[i, 2]/2], [0, pred_size[i, 1]/2, 0], [pred_size[i, 0]/2, 0, 0]]).transpose()
transformed_axes = transform_coordinates_3d(xyz_axis, RT)
projected_axes = calculate_2d_projections(transformed_axes, intrinsics)

[hetolin]

Yeah, this visualization result is right

[taeyeopl]

Great!! Again, Thanks for your quick reply and for sharing the results.