chamfer for empty pointclouds #1174

Summary: Fix divide by zero for empty pointcloud in chamfer. Also for empty batches. In process, needed to regularize num_points_per_cloud for empty batches.

Reviewed By: kjchalup

Differential Revision: D36311330

fbshipit-source-id: 3378ab738bee77ecc286f2110a5c8dc445960340

pytorch3d/implicitron/models/model_dbir.py

@@ -99,7 +99,7 @@ class ModelDBIR(ImplicitronModelBase, torch.nn.Module):
             mask_fg[is_known_idx],
         )
 
-        pcl_size = int(point_cloud.num_points_per_cloud())
+        pcl_size = point_cloud.num_points_per_cloud().item()
         if (self.max_points > 0) and (pcl_size > self.max_points):
             prm = torch.randperm(pcl_size)[: self.max_points]
             point_cloud = Pointclouds(

pytorch3d/loss/chamfer.py

@@ -197,11 +197,13 @@ def chamfer_distance(
         cham_norm_x = cham_norm_x.sum(1)  # (N,)
         cham_norm_y = cham_norm_y.sum(1)  # (N,)
     if point_reduction == "mean":
-        cham_x /= x_lengths
+        x_lengths_clamped = x_lengths.clamp(min=1)
-        cham_y /= y_lengths
+        y_lengths_clamped = y_lengths.clamp(min=1)
+        cham_x /= x_lengths_clamped
+        cham_y /= y_lengths_clamped
         if return_normals:
-            cham_norm_x /= x_lengths
+            cham_norm_x /= x_lengths_clamped
-            cham_norm_y /= y_lengths
+            cham_norm_y /= y_lengths_clamped
 
     if batch_reduction is not None:
         # batch_reduction == "sum"
@@ -211,7 +213,7 @@ def chamfer_distance(
             cham_norm_x = cham_norm_x.sum()
             cham_norm_y = cham_norm_y.sum()
         if batch_reduction == "mean":
-            div = weights.sum() if weights is not None else N
+            div = weights.sum() if weights is not None else max(N, 1)
             cham_x /= div
             cham_y /= div
             if return_normals:

pytorch3d/loss/point_mesh_distance.py

@@ -303,6 +303,7 @@ def point_mesh_edge_distance(meshes: Meshes, pcls: Pointclouds):
     # weight each example by the inverse of number of points in the example
     point_to_cloud_idx = pcls.packed_to_cloud_idx()  # (sum(P_i), )
     num_points_per_cloud = pcls.num_points_per_cloud()  # (N,)
+    # pyre-ignore[16]: `torch.Tensor` has no attribute `gather`
     weights_p = num_points_per_cloud.gather(0, point_to_cloud_idx)
     weights_p = 1.0 / weights_p.float()
     point_to_edge = point_to_edge * weights_p
@@ -377,6 +378,7 @@ def point_mesh_face_distance(
     # weight each example by the inverse of number of points in the example
     point_to_cloud_idx = pcls.packed_to_cloud_idx()  # (sum(P_i),)
     num_points_per_cloud = pcls.num_points_per_cloud()  # (N,)
+    # pyre-ignore[16]: `torch.Tensor` has no attribute `gather`
     weights_p = num_points_per_cloud.gather(0, point_to_cloud_idx)
     weights_p = 1.0 / weights_p.float()
     point_to_face = point_to_face * weights_p

pytorch3d/structures/pointclouds.py

@@ -185,7 +185,6 @@ class Pointclouds:
             self._points_list = points
             self._N = len(self._points_list)
             self.valid = torch.zeros((self._N,), dtype=torch.bool, device=self.device)
-            self._num_points_per_cloud = []
 
             if self._N > 0:
                 self.device = self._points_list[0].device
@@ -208,6 +207,8 @@ class Pointclouds:
                 if len(num_points_per_cloud.unique()) == 1:
                     self.equisized = True
                 self._num_points_per_cloud = num_points_per_cloud
+            else:
+                self._num_points_per_cloud = torch.tensor([], dtype=torch.int64)
 
         elif torch.is_tensor(points):
             if points.dim() != 3 or points.shape[2] != 3:
@@ -525,7 +526,7 @@ class Pointclouds:
         self._compute_packed()
         return self._cloud_to_packed_first_idx
 
-    def num_points_per_cloud(self):
+    def num_points_per_cloud(self) -> torch.Tensor:
         """
         Return a 1D tensor x with length equal to the number of clouds giving
         the number of points in each cloud.

tests/test_chamfer.py

@@ -778,6 +778,17 @@ class TestChamfer(TestCaseMixin, unittest.TestCase):
         with self.assertRaisesRegex(ValueError, "Support for 1 or 2 norm."):
             chamfer_distance(p1, p2, norm=3)
 
+    def test_empty_clouds(self):
+        # Check that point_reduction doesn't divide by zero
+        points1 = Pointclouds(points=[torch.zeros(0, 3), torch.zeros(10, 3)])
+        points2 = Pointclouds(points=torch.ones(2, 40, 3))
+        loss, _ = chamfer_distance(points1, points2, batch_reduction=None)
+        self.assertClose(loss, torch.tensor([0.0, 6.0]))
+
+        # Check that batch_reduction doesn't divide by zero
+        loss2, _ = chamfer_distance(Pointclouds([]), Pointclouds([]))
+        self.assertClose(loss2, torch.tensor(0.0))
+
     @staticmethod
     def chamfer_with_init(
         batch_size: int,