make points2volumes feature rescaling optional

Summary: Add option to not rescale the features, giving more control. https://github.com/facebookresearch/pytorch3d/issues/1137

Reviewed By: nikhilaravi

Differential Revision: D35219577

fbshipit-source-id: cbbb643b91b71bc908cedc6dac0f63f6d1355c85

pytorch3d/ops/points_to_volumes.py

@@ -192,6 +192,7 @@ def add_pointclouds_to_volumes(
     initial_volumes: "Volumes",
     mode: str = "trilinear",
     min_weight: float = 1e-4,
+    rescale_features: bool = True,
     _python: bool = False,
 ) -> "Volumes":
     """
@@ -250,6 +251,10 @@ def add_pointclouds_to_volumes(
         min_weight: A scalar controlling the lowest possible total per-voxel
             weight used to normalize the features accumulated in a voxel.
             Only active for `mode==trilinear`.
+        rescale_features: If False, output features are just the sum of input and
+                            added points. If True, they are averaged. In both cases,
+                            output densities are just summed without rescaling, so
+                            you may need to rescale them afterwards.
         _python: Set to True to use a pure Python implementation, e.g. for test
             purposes, which requires more memory and may be slower.
 
@@ -286,6 +291,7 @@ def add_pointclouds_to_volumes(
         grid_sizes=initial_volumes.get_grid_sizes(),
         mask=mask,
         mode=mode,
+        rescale_features=rescale_features,
         _python=_python,
     )
 
@@ -303,6 +309,7 @@ def add_points_features_to_volume_densities_features(
     min_weight: float = 1e-4,
     mask: Optional[torch.Tensor] = None,
     grid_sizes: Optional[torch.LongTensor] = None,
+    rescale_features: bool = True,
     _python: bool = False,
 ) -> Tuple[torch.Tensor, torch.Tensor]:
     """
@@ -345,6 +352,10 @@ def add_points_features_to_volume_densities_features(
         grid_sizes: `LongTensor` of shape (minibatch, 3) representing the
             spatial resolutions of each of the the non-flattened `volumes` tensors,
             or None to indicate the whole volume is used for every batch element.
+        rescale_features: If False, output features are just the sum of input and
+                            added points. If True, they are averaged. In both cases,
+                            output densities are just summed without rescaling, so
+                            you may need to rescale them afterwards.
         _python: Set to True to use a pure Python implementation.
     Returns:
         volume_features: Output volume of shape `(minibatch, feature_dim, D, H, W)`
@@ -401,12 +412,13 @@ def add_points_features_to_volume_densities_features(
         True,  # align_corners
         splat,
     )
-    if splat:
-        # divide each feature by the total weight of the votes
-        volume_features = volume_features / volume_densities.clamp(min_weight)
-    else:
+
+    if rescale_features:
         # divide each feature by the total weight of the votes
-        volume_features = volume_features / volume_densities.clamp(1.0)
+        if splat:
+            volume_features = volume_features / volume_densities.clamp(min_weight)
+        else:
+            volume_features = volume_features / volume_densities.clamp(1.0)
 
     return volume_features, volume_densities
 

tests/common_testing.py

@@ -192,12 +192,22 @@ class TestCaseMixin(unittest.TestCase):
             self.fail(f"{msg} {err}")
         self.fail(err)
 
-    def assertConstant(self, input: TensorOrArray, value: Real) -> None:
+    def assertConstant(
+        self, input: TensorOrArray, value: Real, *, atol: float = 0
+    ) -> None:
         """
         Asserts input is entirely filled with value.
 
         Args:
             input: tensor or array
+            value: expected value
+            atol: tolerance
         """
-        self.assertEqual(input.min(), value)
-        self.assertEqual(input.max(), value)
+        mn, mx = input.min(), input.max()
+        msg = f"values in range [{mn}, {mx}], not {value}, shape {input.shape}"
+        if atol == 0:
+            self.assertEqual(input.min(), value, msg=msg)
+            self.assertEqual(input.max(), value, msg=msg)
+        else:
+            self.assertGreater(input.min(), value - atol, msg=msg)
+            self.assertLess(input.max(), value + atol, msg=msg)

tests/test_points_to_volumes.py

@@ -387,6 +387,23 @@ class TestPointsToVolumes(TestCaseMixin, unittest.TestCase):
         )
         self.assertClose(torch.sum(densities), torch.tensor(30 * 1000.0), atol=0.1)
 
+    def test_unscaled(self):
+        D = 5
+        P = 1000
+        B, C, H, W = 2, 3, D, D
+        densities = torch.zeros(B, 1, D, H, W)
+        features = torch.zeros(B, C, D, H, W)
+        volumes = Volumes(densities=densities, features=features)
+        points = torch.rand(B, 1000, 3) * (D - 1) - ((D - 1) * 0.5)
+        point_features = torch.rand(B, 1000, C)
+        pointclouds = Pointclouds(points=points, features=point_features)
+
+        volumes2 = add_pointclouds_to_volumes(
+            pointclouds, volumes, rescale_features=False
+        )
+        self.assertConstant(volumes2.densities().sum([2, 3, 4]) / P, 1, atol=1e-5)
+        self.assertConstant(volumes2.features().sum([2, 3, 4]) / P, 0.5, atol=0.03)
+
     def _check_volume_slice_color_density(
         self, V, split_dim, interp_mode, clr_gt, slice_type, border=3
     ):