Join cameras as batch

Summary:
Function to join a list of cameras objects into a single batched object.

FB: In the next diff I will remove the `concatenate_cameras` function in implicitron and update the callsites.

Reviewed By: nikhilaravi

Differential Revision: D33198209

fbshipit-source-id: 0c9f5f5df498a0def9dba756c984e6a946618158

pytorch3d/renderer/__init__.py

@@ -10,7 +10,7 @@ from .blending import (
     sigmoid_alpha_blend,
     softmax_rgb_blend,
 )
-from .camera_utils import rotate_on_spot
+from .camera_utils import join_cameras_as_batch, rotate_on_spot
 from .cameras import OpenGLOrthographicCameras  # deprecated
 from .cameras import OpenGLPerspectiveCameras  # deprecated
 from .cameras import SfMOrthographicCameras  # deprecated
@@ -29,6 +29,7 @@ from .implicit import (
     AbsorptionOnlyRaymarcher,
     EmissionAbsorptionRaymarcher,
     GridRaysampler,
+    HarmonicEmbedding,
     ImplicitRenderer,
     MonteCarloRaysampler,
     NDCGridRaysampler,
@@ -37,7 +38,6 @@ from .implicit import (
     VolumeSampler,
     ray_bundle_to_ray_points,
     ray_bundle_variables_to_ray_points,
-    HarmonicEmbedding,
 )
 from .lighting import AmbientLights, DirectionalLights, PointLights, diffuse, specular
 from .materials import Materials

pytorch3d/renderer/camera_utils.py

@@ -4,11 +4,13 @@
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 
-from typing import Tuple
+from typing import Sequence, Tuple
 
 import torch
 from pytorch3d.transforms import Transform3d
 
+from .cameras import CamerasBase
+
 
 def camera_to_eye_at_up(
     world_to_view_transform: Transform3d,
@@ -141,3 +143,65 @@ def rotate_on_spot(
     new_T = torch.matmul(new_R.transpose(1, 2), old_RT)[:, :, 0]
 
     return new_R, new_T
+
+
+def join_cameras_as_batch(cameras_list: Sequence[CamerasBase]) -> CamerasBase:
+    """
+    Create a batched cameras object by concatenating a list of input
+    cameras objects. All the tensor attributes will be joined along
+    the batch dimension.
+
+    Args:
+        cameras_list: List of camera classes all of the same type and
+            on the same device. Each represents one or more cameras.
+    Returns:
+        cameras: single batched cameras object of the same
+            type as all the objects in the input list.
+    """
+    # Get the type and fields to join from the first camera in the batch
+    c0 = cameras_list[0]
+    fields = c0._FIELDS
+    shared_fields = c0._SHARED_FIELDS
+
+    if not all(isinstance(c, CamerasBase) for c in cameras_list):
+        raise ValueError("cameras in cameras_list must inherit from CamerasBase")
+
+    if not all(type(c) is type(c0) for c in cameras_list[1:]):
+        raise ValueError("All cameras must be of the same type")
+
+    if not all(c.device == c0.device for c in cameras_list[1:]):
+        raise ValueError("All cameras in the batch must be on the same device")
+
+    # Concat the fields to make a batched tensor
+    kwargs = {}
+    kwargs["device"] = c0.device
+
+    for field in fields:
+        field_not_none = [(getattr(c, field) is not None) for c in cameras_list]
+        if not any(field_not_none):
+            continue
+        if not all(field_not_none):
+            raise ValueError(f"Attribute {field} is inconsistently present")
+
+        attrs_list = [getattr(c, field) for c in cameras_list]
+
+        if field in shared_fields:
+            # Only needs to be set once
+            if not all(a == attrs_list[0] for a in attrs_list):
+                raise ValueError(f"Attribute {field} is not constant across inputs")
+
+            # e.g. "in_ndc" is set as attribute "_in_ndc" on the class
+            # but provided as "in_ndc" in the input args
+            if field.startswith("_"):
+                field = field[1:]
+
+            kwargs[field] = attrs_list[0]
+        elif isinstance(attrs_list[0], torch.Tensor):
+            # In the init, all inputs will be converted to
+            # batched tensors before set as attributes
+            # Join as a tensor along the batch dimension
+            kwargs[field] = torch.cat(attrs_list, dim=0)
+        else:
+            raise ValueError(f"Field {field} type is not supported for batching")
+
+    return c0.__class__(**kwargs)

pytorch3d/renderer/cameras.py

@@ -77,7 +77,12 @@ class CamerasBase(TensorProperties):
 
     # Used in __getitem__ to index the relevant fields
     # When creating a new camera, this should be set in the __init__
-    _FIELDS: Tuple = ()
+    _FIELDS: Tuple[str, ...] = ()
+
+    # Names of fields which are a constant property of the whole batch, rather
+    # than themselves a batch of data.
+    # When joining objects into a batch, they will have to agree.
+    _SHARED_FIELDS: Tuple[str, ...] = ()
 
     def get_projection_transform(self):
         """
@@ -499,6 +504,8 @@ class FoVPerspectiveCameras(CamerasBase):
         "degrees",
     )
 
+    _SHARED_FIELDS = ("degrees",)
+
     def __init__(
         self,
         znear=1.0,
@@ -997,6 +1004,8 @@ class PerspectiveCameras(CamerasBase):
         "image_size",
     )
 
+    _SHARED_FIELDS = ("_in_ndc",)
+
     def __init__(
         self,
         focal_length=1.0,
@@ -1047,6 +1056,12 @@ class PerspectiveCameras(CamerasBase):
         else:
             self.image_size = None
 
+        # When focal length is provided as one value, expand to
+        # create (N, 2) shape tensor
+        if self.focal_length.ndim == 1:  # (N,)
+            self.focal_length = self.focal_length[:, None]  # (N, 1)
+        self.focal_length = self.focal_length.expand(-1, 2)  # (N, 2)
+
     def get_projection_transform(self, **kwargs) -> Transform3d:
         """
         Calculate the projection matrix using the
@@ -1227,6 +1242,8 @@ class OrthographicCameras(CamerasBase):
         "image_size",
     )
 
+    _SHARED_FIELDS = ("_in_ndc",)
+
     def __init__(
         self,
         focal_length=1.0,
@@ -1276,6 +1293,12 @@ class OrthographicCameras(CamerasBase):
         else:
             self.image_size = None
 
+        # When focal length is provided as one value, expand to
+        # create (N, 2) shape tensor
+        if self.focal_length.ndim == 1:  # (N,)
+            self.focal_length = self.focal_length[:, None]  # (N, 1)
+        self.focal_length = self.focal_length.expand(-1, 2)  # (N, 2)
+
     def get_projection_transform(self, **kwargs) -> Transform3d:
         """
         Calculate the projection matrix using

tests/test_camera_pixels.py

@@ -250,7 +250,4 @@ class TestPixels(TestCaseMixin, unittest.TestCase):
                             ],
                             dim=1,
                         )
-
-                    print(wanted)
-                    print(camera_points[batch_idx])
                     self.assertClose(camera_points[batch_idx], wanted)

tests/test_cameras.py

@@ -36,6 +36,7 @@ import unittest
 import numpy as np
 import torch
 from common_testing import TestCaseMixin
+from pytorch3d.renderer.camera_utils import join_cameras_as_batch
 from pytorch3d.renderer.cameras import (
     CamerasBase,
     FoVOrthographicCameras,
@@ -688,6 +689,99 @@ class TestCamerasCommon(TestCaseMixin, unittest.TestCase):
                 else:
                     self.assertTrue(val == val_clone)
 
+    def test_join_cameras_as_batch_errors(self):
+        cam0 = PerspectiveCameras(device="cuda:0")
+        cam1 = OrthographicCameras(device="cuda:0")
+
+        # Cameras not of the same type
+        with self.assertRaisesRegex(ValueError, "same type"):
+            join_cameras_as_batch([cam0, cam1])
+
+        cam2 = OrthographicCameras(device="cpu")
+        # Cameras not on the same device
+        with self.assertRaisesRegex(ValueError, "same device"):
+            join_cameras_as_batch([cam1, cam2])
+
+        cam3 = OrthographicCameras(in_ndc=False, device="cuda:0")
+        # Different coordinate systems -- all should be in ndc or in screen
+        with self.assertRaisesRegex(
+            ValueError, "Attribute _in_ndc is not constant across inputs"
+        ):
+            join_cameras_as_batch([cam1, cam3])
+
+    def join_cameras_as_batch_fov(self, camera_cls):
+        R0 = torch.randn((6, 3, 3))
+        R1 = torch.randn((3, 3, 3))
+        cam0 = camera_cls(znear=10.0, zfar=100.0, R=R0, device="cuda:0")
+        cam1 = camera_cls(znear=10.0, zfar=200.0, R=R1, device="cuda:0")
+
+        cam_batch = join_cameras_as_batch([cam0, cam1])
+
+        self.assertEqual(cam_batch._N, cam0._N + cam1._N)
+        self.assertEqual(cam_batch.device, cam0.device)
+        self.assertClose(cam_batch.R, torch.cat((R0, R1), dim=0).to(device="cuda:0"))
+
+    def join_cameras_as_batch(self, camera_cls):
+        R0 = torch.randn((6, 3, 3))
+        R1 = torch.randn((3, 3, 3))
+        p0 = torch.randn((6, 2, 1))
+        p1 = torch.randn((3, 2, 1))
+        f0 = 5.0
+        f1 = torch.randn(3, 2)
+        f2 = torch.randn(3, 1)
+        cam0 = camera_cls(
+            R=R0,
+            focal_length=f0,
+            principal_point=p0,
+        )
+        cam1 = camera_cls(
+            R=R1,
+            focal_length=f0,
+            principal_point=p1,
+        )
+        cam2 = camera_cls(
+            R=R1,
+            focal_length=f1,
+            principal_point=p1,
+        )
+        cam3 = camera_cls(
+            R=R1,
+            focal_length=f2,
+            principal_point=p1,
+        )
+        cam_batch = join_cameras_as_batch([cam0, cam1])
+
+        self.assertEqual(cam_batch._N, cam0._N + cam1._N)
+        self.assertEqual(cam_batch.device, cam0.device)
+        self.assertClose(cam_batch.R, torch.cat((R0, R1), dim=0))
+        self.assertClose(cam_batch.principal_point, torch.cat((p0, p1), dim=0))
+        self.assertEqual(cam_batch._in_ndc, cam0._in_ndc)
+
+        # Test one broadcasted value and one fixed value
+        # Focal length as (N,) in one camera and (N, 2) in the other
+        cam_batch = join_cameras_as_batch([cam0, cam2])
+        self.assertEqual(cam_batch._N, cam0._N + cam2._N)
+        self.assertClose(cam_batch.R, torch.cat((R0, R1), dim=0))
+        self.assertClose(
+            cam_batch.focal_length,
+            torch.cat([torch.tensor([[f0, f0]]).expand(6, -1), f1], dim=0),
+        )
+
+        # Focal length as (N, 1) in one camera and (N, 2) in the other
+        cam_batch = join_cameras_as_batch([cam2, cam3])
+        self.assertClose(
+            cam_batch.focal_length,
+            torch.cat([f1, f2.expand(-1, 2)], dim=0),
+        )
+
+    def test_join_batch_perspective(self):
+        self.join_cameras_as_batch_fov(FoVPerspectiveCameras)
+        self.join_cameras_as_batch(PerspectiveCameras)
+
+    def test_join_batch_orthographic(self):
+        self.join_cameras_as_batch_fov(FoVOrthographicCameras)
+        self.join_cameras_as_batch(OrthographicCameras)
+
 
 ############################################################
 #                FoVPerspective Camera                     #
@@ -1055,7 +1149,7 @@ class TestOrthographicProjection(TestCaseMixin, unittest.TestCase):
         index = torch.tensor([1, 3, 5], dtype=torch.int64)
         c135 = cam[index]
         self.assertEqual(len(c135), 3)
-        self.assertClose(c135.focal_length, torch.tensor([5.0] * 3))
+        self.assertClose(c135.focal_length, torch.tensor([[5.0, 5.0]] * 3))
         self.assertClose(c135.R, R_matrix[[1, 3, 5], ...])
         self.assertClose(c135.principal_point, principal_point[[1, 3, 5], ...])
 
@@ -1131,7 +1225,7 @@ class TestPerspectiveProjection(TestCaseMixin, unittest.TestCase):
         index = torch.tensor([1, 3, 5], dtype=torch.int64)
         c135 = cam[index]
         self.assertEqual(len(c135), 3)
-        self.assertClose(c135.focal_length, torch.tensor([5.0] * 3))
+        self.assertClose(c135.focal_length, torch.tensor([[5.0, 5.0]] * 3))
         self.assertClose(c135.R, R_matrix[[1, 3, 5], ...])
         self.assertClose(c135.principal_point, principal_point[[1, 3, 5], ...])