Updates to cameras and rasterizer to infer camera type correctly

Summary: Small update to the cameras and rasterizer to correctly infer the type of camera (perspective vs orthographic).

Reviewed By: jcjohnson

Differential Revision: D26267225

fbshipit-source-id: a58ed3bc2ab25553d2a4307c734204c1d41b5176

pytorch3d/renderer/cameras.py

@@ -42,15 +42,16 @@ class CamerasBase(TensorProperties):
 
     It defines methods that are common to all camera models:
         - `get_camera_center` that returns the optical center of the camera in
-    world coordinates
+            world coordinates
         - `get_world_to_view_transform` which returns a 3D transform from
-    world coordinates to the camera view coordinates (R, T)
+            world coordinates to the camera view coordinates (R, T)
         - `get_full_projection_transform` which composes the projection
-    transform (P) with the world-to-view transform (R, T)
+            transform (P) with the world-to-view transform (R, T)
         - `transform_points` which takes a set of input points in world coordinates and
-    projects to NDC coordinates ranging from [-1, -1, znear] to [+1, +1, zfar].
+            projects to NDC coordinates ranging from [-1, -1, znear] to [+1, +1, zfar].
         - `transform_points_screen` which takes a set of input points in world coordinates and
-    projects them to the screen coordinates ranging from [0, 0, znear] to [W-1, H-1, zfar]
+            projects them to the screen coordinates ranging from
+            [0, 0, znear] to [W-1, H-1, zfar]
 
     For each new camera, one should implement the `get_projection_transform`
     routine that returns the mapping from camera view coordinates to NDC coordinates.
@@ -268,6 +269,12 @@ class CamerasBase(TensorProperties):
         other = cam_type(device=self.device)
         return super().clone(other)
 
+    def is_perspective(self):
+        raise NotImplementedError()
+
+    def get_znear(self):
+        return self.znear if hasattr(self, "znear") else None
+
 
 ############################################################
 #             Field of View Camera Classes                 #
@@ -534,6 +541,9 @@ class FoVPerspectiveCameras(CamerasBase):
         unprojection_transform = to_ndc_transform.inverse()
         return unprojection_transform.transform_points(xy_sdepth)
 
+    def is_perspective(self):
+        return True
+
 
 def OpenGLOrthographicCameras(
     znear=1.0,
@@ -752,6 +762,9 @@ class FoVOrthographicCameras(CamerasBase):
         unprojection_transform = to_ndc_transform.inverse()
         return unprojection_transform.transform_points(xy_sdepth)
 
+    def is_perspective(self):
+        return False
+
 
 ############################################################
 #             MultiView Camera Classes                     #
@@ -927,6 +940,9 @@ class PerspectiveCameras(CamerasBase):
         )
         return unprojection_transform.transform_points(xy_inv_depth)
 
+    def is_perspective(self):
+        return True
+
 
 def SfMOrthographicCameras(
     focal_length=1.0, principal_point=((0.0, 0.0),), R=_R, T=_T, device="cpu"
@@ -1086,6 +1102,9 @@ class OrthographicCameras(CamerasBase):
         unprojection_transform = to_ndc_transform.inverse()
         return unprojection_transform.transform_points(xy_depth)
 
+    def is_perspective(self):
+        return False
+
 
 ################################################
 #       Helper functions for cameras           #

pytorch3d/renderer/mesh/rasterizer.py

@@ -139,8 +139,13 @@ class MeshRasterizer(nn.Module):
         if clip_barycentric_coords is None:
             clip_barycentric_coords = raster_settings.blur_radius > 0.0
 
-        # TODO(jcjohns): Should we try to set perspective_correct automatically
-        # based on the type of the camera?
+        # If not specified, infer perspective_correct from the camera
+        cameras = kwargs.get("cameras", self.cameras)
+        if raster_settings.perspective_correct is not None:
+            perspective_correct = raster_settings.perspective_correct
+        else:
+            perspective_correct = cameras.is_perspective()
+
         pix_to_face, zbuf, bary_coords, dists = rasterize_meshes(
             meshes_screen,
             image_size=raster_settings.image_size,
@@ -148,7 +153,7 @@ class MeshRasterizer(nn.Module):
             faces_per_pixel=raster_settings.faces_per_pixel,
             bin_size=raster_settings.bin_size,
             max_faces_per_bin=raster_settings.max_faces_per_bin,
-            perspective_correct=raster_settings.perspective_correct,
+            perspective_correct=perspective_correct,
             clip_barycentric_coords=clip_barycentric_coords,
             cull_backfaces=raster_settings.cull_backfaces,
         )