6D representation of rotations.

Summary: Conversion to/from the 6D representation of rotation from the paper http://arxiv.org/abs/1812.07035 ; based on David’s implementation.

Reviewed By: davnov134

Differential Revision: D22234397

fbshipit-source-id: 9e25ee93da7e3a2f2068cbe362cb5edc88649ce0

pytorch3d/transforms/rotation_conversions.py

@@ -4,6 +4,7 @@ import functools
 from typing import Optional
 
 import torch
+import torch.nn.functional as F
 
 
 """
@@ -252,7 +253,7 @@ def random_quaternions(
     i.e. versors with nonnegative real part.
 
     Args:
-        n: Number to return.
+        n: Number of quaternions in a batch to return.
         dtype: Type to return.
         device: Desired device of returned tensor. Default:
             uses the current device for the default tensor type.
@@ -275,7 +276,7 @@ def random_rotations(
     Generate random rotations as 3x3 rotation matrices.
 
     Args:
-        n: Number to return.
+        n: Number of rotation matrices in a batch to return.
         dtype: Type to return.
         device: Device of returned tensor. Default: if None,
             uses the current device for the default tensor type.
@@ -400,3 +401,45 @@ def quaternion_apply(quaternion, point):
         quaternion_invert(quaternion),
     )
     return out[..., 1:]
+
+
+def rotation_6d_to_matrix(d6: torch.Tensor) -> torch.Tensor:
+    """
+    Converts 6D rotation representation by Zhou et al. [1] to rotation matrix
+    using Gram--Schmidt orthogonalisation per Section B of [1].
+    Args:
+        d6: 6D rotation representation, of size (*, 6)
+
+    Returns:
+        batch of rotation matrices of size (*, 3, 3)
+
+    [1] Zhou, Y., Barnes, C., Lu, J., Yang, J., & Li, H.
+    On the Continuity of Rotation Representations in Neural Networks.
+    IEEE Conference on Computer Vision and Pattern Recognition, 2019.
+    Retrieved from http://arxiv.org/abs/1812.07035
+    """
+
+    a1, a2 = d6[..., :3], d6[..., 3:]
+    b1 = F.normalize(a1, dim=-1)
+    b2 = a2 - (b1 * a2).sum(-1, keepdim=True) * b1
+    b2 = F.normalize(b2, dim=-1)
+    b3 = torch.cross(b1, b2, dim=-1)
+    return torch.stack((b1, b2, b3), dim=-2)
+
+
+def matrix_to_rotation_6d(matrix: torch.Tensor) -> torch.Tensor:
+    """
+    Converts rotation matrices to 6D rotation representation by Zhou et al. [1]
+    by dropping the last row. Note that 6D representation is not unique.
+    Args:
+        matrix: batch of rotation matrices of size (*, 3, 3)
+
+    Returns:
+        6D rotation representation, of size (*, 6)
+
+    [1] Zhou, Y., Barnes, C., Lu, J., Yang, J., & Li, H.
+    On the Continuity of Rotation Representations in Neural Networks.
+    IEEE Conference on Computer Vision and Pattern Recognition, 2019.
+    Retrieved from http://arxiv.org/abs/1812.07035
+    """
+    return matrix[..., :2, :].clone().reshape(*matrix.size()[:-2], 6)

pytorch3d/transforms/so3.py

@@ -10,7 +10,7 @@ HAT_INV_SKEW_SYMMETRIC_TOL = 1e-5
 def so3_relative_angle(R1, R2, cos_angle: bool = False):
     """
     Calculates the relative angle (in radians) between pairs of
-    rotation matrices `R1` and `R2` with `angle = acos(0.5 * Trace(R1 R2^T)-1)`
+    rotation matrices `R1` and `R2` with `angle = acos(0.5 * (Trace(R1 R2^T)-1))`
 
     .. note::
         This corresponds to a geodesic distance on the 3D manifold of rotation

pytorch3d/transforms/transform3d.py

@@ -45,9 +45,9 @@ class Transform3d:
 
     .. code-block:: python
 
-        y1 = t3.transform_points(t2.transform_points(t2.transform_points(x)))
-        y2 = t1.compose(t2).compose(t3).transform_points()
-        y3 = t1.compose(t2, t3).transform_points()
+        y1 = t3.transform_points(t2.transform_points(t1.transform_points(x)))
+        y2 = t1.compose(t2).compose(t3).transform_points(x)
+        y3 = t1.compose(t2, t3).transform_points(x)
 
 
     Composing transforms should broadcast.

tests/test_rotation_conversions.py

@@ -6,16 +6,19 @@ import math
 import unittest
 
 import torch
+from common_testing import TestCaseMixin
 from pytorch3d.transforms.rotation_conversions import (
     euler_angles_to_matrix,
     matrix_to_euler_angles,
     matrix_to_quaternion,
+    matrix_to_rotation_6d,
     quaternion_apply,
     quaternion_multiply,
     quaternion_to_matrix,
     random_quaternions,
     random_rotation,
     random_rotations,
+    rotation_6d_to_matrix,
 )
 
 
@@ -48,7 +51,7 @@ class TestRandomRotation(unittest.TestCase):
             self.assertLess(chisquare_statistic, 12, (counts, chisquare_statistic, k))
 
 
-class TestRotationConversion(unittest.TestCase):
+class TestRotationConversion(TestCaseMixin, unittest.TestCase):
     def setUp(self) -> None:
         super().setUp()
         torch.manual_seed(1)
@@ -154,3 +157,31 @@ class TestRotationConversion(unittest.TestCase):
         [p, q] = torch.autograd.grad(transform1.sum(), [points, quaternions])
         self.assertTrue(torch.isfinite(p).all())
         self.assertTrue(torch.isfinite(q).all())
+
+    def test_6d(self):
+        """Converting to 6d and back"""
+        r = random_rotations(13, dtype=torch.float64)
+
+        # 6D representation is not unique,
+        # but we implement it by taking the first two rows of the matrix
+        r6d = matrix_to_rotation_6d(r)
+        self.assertClose(r6d, r[:, :2, :].reshape(-1, 6))
+
+        # going to 6D and back should not change the matrix
+        r_hat = rotation_6d_to_matrix(r6d)
+        self.assertClose(r_hat, r)
+
+        # moving the second row R2 in the span of (R1, R2) should not matter
+        r6d[:, 3:] += 2 * r6d[:, :3]
+        r6d[:, :3] *= 3.0
+        r_hat = rotation_6d_to_matrix(r6d)
+        self.assertClose(r_hat, r)
+
+        # check that we map anything to a valid rotation
+        r6d = torch.rand(13, 6)
+        r6d[:4, :] *= 3.0
+        r6d[4:8, :] -= 0.5
+        r = rotation_6d_to_matrix(r6d)
+        self.assertClose(
+            torch.matmul(r, r.permute(0, 2, 1)), torch.eye(3).expand_as(r), atol=1e-6
+        )