The purpose of base in loss.py

[tdt98]

@j96w , thank you for sharing your great work. I have read your code and i have problem in understanding those lines of code:

base = torch.cat(((1.0 - 2.0*(pred_r[:, :, 2]**2 + pred_r[:, :, 3]**2)).view(bs, num_p, 1),\
                      (2.0*pred_r[:, :, 1]*pred_r[:, :, 2] - 2.0*pred_r[:, :, 0]*pred_r[:, :, 3]).view(bs, num_p, 1), \
                      (2.0*pred_r[:, :, 0]*pred_r[:, :, 2] + 2.0*pred_r[:, :, 1]*pred_r[:, :, 3]).view(bs, num_p, 1), \
                      (2.0*pred_r[:, :, 1]*pred_r[:, :, 2] + 2.0*pred_r[:, :, 3]*pred_r[:, :, 0]).view(bs, num_p, 1), \
                      (1.0 - 2.0*(pred_r[:, :, 1]**2 + pred_r[:, :, 3]**2)).view(bs, num_p, 1), \
                      (-2.0*pred_r[:, :, 0]*pred_r[:, :, 1] + 2.0*pred_r[:, :, 2]*pred_r[:, :, 3]).view(bs, num_p, 1), \
                      (-2.0*pred_r[:, :, 0]*pred_r[:, :, 2] + 2.0*pred_r[:, :, 1]*pred_r[:, :, 3]).view(bs, num_p, 1), \
                      (2.0*pred_r[:, :, 0]*pred_r[:, :, 1] + 2.0*pred_r[:, :, 2]*pred_r[:, :, 3]).view(bs, num_p, 1), \
                      (1.0 - 2.0*(pred_r[:, :, 1]**2 + pred_r[:, :, 2]**2)).view(bs, num_p, 1)), dim=2).contiguous().view(bs * num_p, 3, 3)
    ori_base = base
    base = base.contiguous().transpose(2, 1).contiguous()
    model_points = model_points.view(bs, 1, num_point_mesh, 3).repeat(1, num_p, 1, 1).view(bs * num_p, num_point_mesh, 3)
    target = target.view(bs, 1, num_point_mesh, 3).repeat(1, num_p, 1, 1).view(bs * num_p, num_point_mesh, 3)
    ori_target = target
    pred_t = pred_t.contiguous().view(bs * num_p, 1, 3)
    ori_t = pred_t
    points = points.contiguous().view(bs * num_p, 1, 3)
    pred_c = pred_c.contiguous().view(bs * num_p)

    pred = torch.add(torch.bmm(model_points, base), points + pred_t)

Why do we need to transform the pred_r into base? Could we calculate the pred as normal?

[Qind1609]

I also have problem in understanding those lines, Can someone please explain it?