add from_ndc to unproject_points

bottler · facebook-github-bot · commit cff4876131e7 · 2021-12-07T15:04:20.000-08:00
Summary: Give unproject_points an argument letting it bypass screen space. use it to let the raysampler work for cameras defined in screen space.

Reviewed By: gkioxari

Differential Revision: D32596600

fbshipit-source-id: 2fe585dcd138cdbc65dd1c70e1957fd894512d3d
diff --git a/pytorch3d/renderer/cameras.py b/pytorch3d/renderer/cameras.py
@@ -129,6 +129,10 @@ def unproject_points(self):
                 coordinates using the camera extrinsics `R` and `T`.
                 `False` ignores `R` and `T` and unprojects to
                 the camera view coordinates.
+            from_ndc: If `False` (default), assumes xy part of input is in
+                NDC space if self.in_ndc(), otherwise in screen space. If
+                `True`, assumes xy is in NDC space even if the camera
+                is defined in screen space.
 
         Returns
             new_points: unprojected points with the same shape as `xy_depth`.
@@ -998,12 +1002,27 @@ def get_projection_transform(self, **kwargs) -> Transform3d:
         return transform
 
     def unproject_points(
-        self, xy_depth: torch.Tensor, world_coordinates: bool = True, **kwargs
+        self,
+        xy_depth: torch.Tensor,
+        world_coordinates: bool = True,
+        from_ndc: bool = False,
+        **kwargs
     ) -> torch.Tensor:
+        """
+        Args:
+            from_ndc: If `False` (default), assumes xy part of input is in
+                NDC space if self.in_ndc(), otherwise in screen space. If
+                `True`, assumes xy is in NDC space even if the camera
+                is defined in screen space.
+        """
         if world_coordinates:
             to_camera_transform = self.get_full_projection_transform(**kwargs)
         else:
             to_camera_transform = self.get_projection_transform(**kwargs)
+        if from_ndc:
+            to_camera_transform = to_camera_transform.compose(
+                self.get_ndc_camera_transform()
+            )
 
         unprojection_transform = to_camera_transform.inverse()
         xy_inv_depth = torch.cat(
@@ -1030,6 +1049,8 @@ def get_ndc_camera_transform(self, **kwargs) -> Transform3d:
         For cameras defined in screen space, we adjust the principal point computation
         which is defined in the image space (commonly) and scale the points to NDC space.
 
+        This transform leaves the depth unchanged.
+
         Important: This transforms assumes PyTorch3D conventions for the input points,
         i.e. +X left, +Y up.
         """
@@ -1199,12 +1220,27 @@ def get_projection_transform(self, **kwargs) -> Transform3d:
         return transform
 
     def unproject_points(
-        self, xy_depth: torch.Tensor, world_coordinates: bool = True, **kwargs
+        self,
+        xy_depth: torch.Tensor,
+        world_coordinates: bool = True,
+        from_ndc: bool = False,
+        **kwargs
     ) -> torch.Tensor:
+        """
+        Args:
+            from_ndc: If `False` (default), assumes xy part of input is in
+                NDC space if self.in_ndc(), otherwise in screen space. If
+                `True`, assumes xy is in NDC space even if the camera
+                is defined in screen space.
+        """
         if world_coordinates:
             to_camera_transform = self.get_full_projection_transform(**kwargs)
         else:
             to_camera_transform = self.get_projection_transform(**kwargs)
+        if from_ndc:
+            to_camera_transform = to_camera_transform.compose(
+                self.get_ndc_camera_transform()
+            )
 
         unprojection_transform = to_camera_transform.inverse()
         return unprojection_transform.transform_points(xy_depth)
diff --git a/pytorch3d/renderer/implicit/raysampling.py b/pytorch3d/renderer/implicit/raysampling.py
@@ -312,7 +312,7 @@ def _xy_to_ray_bundle(
     )
 
     # unproject the points
-    unprojected = cameras.unproject_points(to_unproject)  # pyre-ignore
+    unprojected = cameras.unproject_points(to_unproject, from_ndc=True)  # pyre-ignore
 
     # split the two planes back
     rays_plane_1_world = unprojected[:, :n_rays_per_image]

Original file line number	Diff line number	Diff line change
`@@ -312,7 +312,7 @@ def _xy_to_ray_bundle(`
`312`	`312`	`)`
`313`	`313`
`314`	`314`	`# unproject the points`
`315`		`- unprojected = cameras.unproject_points(to_unproject) # pyre-ignore`
	`315`	`+ unprojected = cameras.unproject_points(to_unproject, from_ndc=True) # pyre-ignore`
`316`	`316`
`317`	`317`	`# split the two planes back`
`318`	`318`	`rays_plane_1_world = unprojected[:, :n_rays_per_image]`