Clarifies Jacobian calculation in symmetric ICP

Clarifies that the Jacobian calculation in the symmetric ICP
implementation uses target normal for the first 6 elements and
source normal for the last 6 elements.

This is reflected in the documentation and test cases.

Author: eclipse0922

cpp/open3d/t/pipelines/kernel/RegistrationImpl.h

@@ -240,8 +240,8 @@ template bool GetJacobianPointToPlane(int64_t workload_idx,
 /// \param target_normals_ptr Pointer to target point normals (3N elements).
 /// \param correspondence_indices Pointer to correspondence indices.
 /// \param J_ij Output array for Jacobians (must have space for 12 elements:
-///             elements 0-5 for source-to-target Jacobian,
-///             elements 6-11 for target-to-source Jacobian).
+///             elements 0-5 for Jacobian using target normal,
+///             elements 6-11 for Jacobian using source normal).
 /// \param r1 Output residual for source-to-target point-to-plane distance.
 /// \param r2 Output residual for target-to-source point-to-plane distance.
 /// \return true if correspondence is valid, false if correspondence is -1.

docs/jupyter/pipelines/icp_registration.ipynb

@@ -1,10 +1,5 @@
 {
  "cells": [
-  {
-   "cell_type": "markdown",
-   "source": "Symmetric ICP achieves tight alignment comparable to point-to-plane ICP (a `fitness` score around 0.62 and an `inlier_rmse` around 0.0065).\n\n<div class=\"alert alert-info\">\n\n**Note:**\n\nSymmetric ICP requires both source and target point clouds to have normals. The symmetric objective function uses normals from both point clouds to constrain the transformation, which can provide more robust alignment when both sets of normals are reliable.\n\n</div>",
-   "metadata": {}
-  },
   {
    "cell_type": "code",
    "execution_count": null,
@@ -251,6 +246,11 @@
    "metadata": {},
    "execution_count": null,
    "outputs": []
+  },
+  {
+   "cell_type": "markdown",
+   "source": "Symmetric ICP achieves tight alignment comparable to point-to-plane ICP (a `fitness` score around 0.62 and an `inlier_rmse` around 0.0065).\n\n<div class=\"alert alert-info\">\n\n**Note:**\n\nSymmetric ICP requires both source and target point clouds to have normals. The symmetric objective function uses normals from both point clouds to constrain the transformation, which can provide more robust alignment when both sets of normals are reliable.\n\n</div>",
+   "metadata": {}
   }
  ],
  "metadata": {

python/test/test_symmetric_icp.py

@@ -227,7 +227,6 @@ def test_registration_symmetric_icp_convergence(self):
 
         # Create a more complex test case
         source = o3d.geometry.PointCloud()
-        target = o3d.geometry.PointCloud()
 
         # Generate random points with normals
         num_points = 50