Example cleanup

docs/source/examples/07_record3d_visualizer.rst

@@ -18,10 +18,11 @@ Parse and stream record3d captures. To get the demo data, see ``./assets/downloa
 
         import numpy as np
         import tyro
+        from tqdm.auto import tqdm
+
         import viser
         import viser.extras
         import viser.transforms as tf
-        from tqdm.auto import tqdm
 
 
         def main(

docs/source/examples/08_smpl_visualizer.rst

@@ -24,7 +24,6 @@ See here for download instructions:
 
         import numpy as np
         import tyro
-
         import viser
         import viser.transforms as tf
 
@@ -207,12 +206,9 @@ See here for download instructions:
 
                 @gui_random_joints.on_click
                 def _(_):
+                    rng = np.random.default_rng()
                     for joint in gui_joints:
-                        # It's hard to uniformly sample orientations directly in so(3), so we
-                        # first sample on S^3 and then convert.
-                        quat = np.random.normal(loc=0.0, scale=1.0, size=(4,))
-                        quat /= np.linalg.norm(quat)
-                        joint.value = tf.SO3(wxyz=quat).log()
+                        joint.value = tf.SO3.sample_uniform(rng).log()
 
                 gui_joints: list[viser.GuiInputHandle[tuple[float, float, float]]] = []
                 for i in range(num_joints):

docs/source/examples/25_smpl_visualizer_skinned.rst

@@ -25,21 +25,18 @@ See here for download instructions:
 
         import numpy as np
         import tyro
-
         import viser
         import viser.transforms as tf
 
 
         @dataclass(frozen=True)
-        class SmplOutputs:
-            vertices: np.ndarray
-            faces: np.ndarray
+        class SmplFkOutputs:
             T_world_joint: np.ndarray  # (num_joints, 4, 4)
             T_parent_joint: np.ndarray  # (num_joints, 4, 4)
 
 
         class SmplHelper:
-            """Helper for models in the SMPL family, implemented in numpy."""
+            """Helper for models in the SMPL family, implemented in numpy. Does not include blend skinning."""
 
             def __init__(self, model_path: Path) -> None:
                 assert model_path.suffix.lower() == ".npz", "Model should be an .npz file!"
@@ -62,7 +59,9 @@ See here for download instructions:
                 j_tpose = np.einsum("jv,vx->jx", self.J_regressor, v_tpose)
                 return v_tpose, j_tpose
 
-            def get_outputs(self, betas: np.ndarray, joint_rotmats: np.ndarray) -> SmplOutputs:
+            def get_outputs(
+                self, betas: np.ndarray, joint_rotmats: np.ndarray
+            ) -> SmplFkOutputs:
                 # Get shaped vertices + joint positions, when all local poses are identity.
                 v_tpose = self.v_template + np.einsum("vxb,b->vx", self.shapedirs, betas)
                 j_tpose = np.einsum("jv,vx->jx", self.J_regressor, v_tpose)
@@ -78,15 +77,7 @@ See here for download instructions:
                 for i in range(1, self.num_joints):
                     T_world_joint[i] = T_world_joint[self.parent_idx[i]] @ T_parent_joint[i]
 
-                # Linear blend skinning.
-                pose_delta = (joint_rotmats[1:, ...] - np.eye(3)).flatten()
-                v_blend = v_tpose + np.einsum("byn,n->by", self.posedirs, pose_delta)
-                v_delta = np.ones((v_blend.shape[0], self.num_joints, 4))
-                v_delta[:, :, :3] = v_blend[:, None, :] - j_tpose[None, :, :]
-                v_posed = np.einsum(
-                    "jxy,vj,vjy->vx", T_world_joint[:, :3, :], self.weights, v_delta
-                )
-                return SmplOutputs(v_posed, self.faces, T_world_joint, T_parent_joint)
+                return SmplFkOutputs(T_world_joint, T_parent_joint)
 
 
         def main(model_path: Path) -> None:
@@ -252,12 +243,9 @@ See here for download instructions:
 
                 @gui_random_joints.on_click
                 def _(_):
+                    rng = np.random.default_rng()
                     for joint in gui_joints:
-                        # It's hard to uniformly sample orientations directly in so(3), so we
-                        # first sample on S^3 and then convert.
-                        quat = np.random.normal(loc=0.0, scale=1.0, size=(4,))
-                        quat /= np.linalg.norm(quat)
-                        joint.value = tf.SO3(wxyz=quat).log()
+                        joint.value = tf.SO3.sample_uniform(rng).log()
 
                 gui_joints: List[viser.GuiInputHandle[Tuple[float, float, float]]] = []
                 for i in range(num_joints):

examples/08_smpl_visualizer.py

@@ -197,12 +197,9 @@ def _(_):
 
         @gui_random_joints.on_click
         def _(_):
+            rng = np.random.default_rng()
             for joint in gui_joints:
-                # It's hard to uniformly sample orientations directly in so(3), so we
-                # first sample on S^3 and then convert.
-                quat = np.random.normal(loc=0.0, scale=1.0, size=(4,))
-                quat /= np.linalg.norm(quat)
-                joint.value = tf.SO3(wxyz=quat).log()
+                joint.value = tf.SO3.sample_uniform(rng).log()
 
         gui_joints: list[viser.GuiInputHandle[tuple[float, float, float]]] = []
         for i in range(num_joints):

examples/25_smpl_visualizer_skinned.py

@@ -26,15 +26,13 @@
 
 
 @dataclass(frozen=True)
-class SmplOutputs:
-    vertices: np.ndarray
-    faces: np.ndarray
+class SmplFkOutputs:
     T_world_joint: np.ndarray  # (num_joints, 4, 4)
     T_parent_joint: np.ndarray  # (num_joints, 4, 4)
 
 
 class SmplHelper:
-    """Helper for models in the SMPL family, implemented in numpy."""
+    """Helper for models in the SMPL family, implemented in numpy. Does not include blend skinning."""
 
     def __init__(self, model_path: Path) -> None:
         assert model_path.suffix.lower() == ".npz", "Model should be an .npz file!"
@@ -57,7 +55,9 @@ def get_tpose(self, betas: np.ndarray) -> tuple[np.ndarray, np.ndarray]:
         j_tpose = np.einsum("jv,vx->jx", self.J_regressor, v_tpose)
         return v_tpose, j_tpose
 
-    def get_outputs(self, betas: np.ndarray, joint_rotmats: np.ndarray) -> SmplOutputs:
+    def get_outputs(
+        self, betas: np.ndarray, joint_rotmats: np.ndarray
+    ) -> SmplFkOutputs:
         # Get shaped vertices + joint positions, when all local poses are identity.
         v_tpose = self.v_template + np.einsum("vxb,b->vx", self.shapedirs, betas)
         j_tpose = np.einsum("jv,vx->jx", self.J_regressor, v_tpose)
@@ -73,15 +73,7 @@ def get_outputs(self, betas: np.ndarray, joint_rotmats: np.ndarray) -> SmplOutpu
         for i in range(1, self.num_joints):
             T_world_joint[i] = T_world_joint[self.parent_idx[i]] @ T_parent_joint[i]
 
-        # Linear blend skinning.
-        pose_delta = (joint_rotmats[1:, ...] - np.eye(3)).flatten()
-        v_blend = v_tpose + np.einsum("byn,n->by", self.posedirs, pose_delta)
-        v_delta = np.ones((v_blend.shape[0], self.num_joints, 4))
-        v_delta[:, :, :3] = v_blend[:, None, :] - j_tpose[None, :, :]
-        v_posed = np.einsum(
-            "jxy,vj,vjy->vx", T_world_joint[:, :3, :], self.weights, v_delta
-        )
-        return SmplOutputs(v_posed, self.faces, T_world_joint, T_parent_joint)
+        return SmplFkOutputs(T_world_joint, T_parent_joint)
 
 
 def main(model_path: Path) -> None:
@@ -247,12 +239,9 @@ def _(_):
 
         @gui_random_joints.on_click
         def _(_):
+            rng = np.random.default_rng()
             for joint in gui_joints:
-                # It's hard to uniformly sample orientations directly in so(3), so we
-                # first sample on S^3 and then convert.
-                quat = np.random.normal(loc=0.0, scale=1.0, size=(4,))
-                quat /= np.linalg.norm(quat)
-                joint.value = tf.SO3(wxyz=quat).log()
+                joint.value = tf.SO3.sample_uniform(rng).log()
 
         gui_joints: List[viser.GuiInputHandle[Tuple[float, float, float]]] = []
         for i in range(num_joints):