Open3D (C++ API)  0.17.0
FillInLinearSystemImpl.h
Go to the documentation of this file.
1 // ----------------------------------------------------------------------------
2 // - Open3D: www.open3d.org -
3 // ----------------------------------------------------------------------------
4 // Copyright (c) 2018-2023 www.open3d.org
5 // SPDX-License-Identifier: MIT
6 // ----------------------------------------------------------------------------
7 
11 
12 namespace open3d {
13 namespace t {
14 namespace pipelines {
15 namespace kernel {
16 #if defined(__CUDACC__)
17 void FillInRigidAlignmentTermCUDA
18 #else
20 #endif
21  (core::Tensor &AtA,
22  core::Tensor &Atb,
23  core::Tensor &residual,
24  const core::Tensor &Ti_ps,
25  const core::Tensor &Tj_qs,
26  const core::Tensor &Ri_normal_ps,
27  int i,
28  int j,
29  float threshold) {
30 
31  core::Device device = AtA.GetDevice();
32  int64_t n = Ti_ps.GetLength();
33  if (Tj_qs.GetLength() != n || Ri_normal_ps.GetLength() != n) {
35  "Unable to setup linear system: input length mismatch.");
36  }
37 
38  // First fill in a small 12 x 12 linear system
39  core::Tensor AtA_local =
40  core::Tensor::Zeros({12, 12}, core::Float32, device);
41  core::Tensor Atb_local = core::Tensor::Zeros({12}, core::Float32, device);
42 
43  float *AtA_local_ptr = static_cast<float *>(AtA_local.GetDataPtr());
44  float *Atb_local_ptr = static_cast<float *>(Atb_local.GetDataPtr());
45  float *residual_ptr = static_cast<float *>(residual.GetDataPtr());
46 
47  const float *Ti_ps_ptr = static_cast<const float *>(Ti_ps.GetDataPtr());
48  const float *Tj_qs_ptr = static_cast<const float *>(Tj_qs.GetDataPtr());
49  const float *Ri_normal_ps_ptr =
50  static_cast<const float *>(Ri_normal_ps.GetDataPtr());
51 
53  AtA.GetDevice(), n, [=] OPEN3D_DEVICE(int64_t workload_idx) {
54  const float *p_prime = Ti_ps_ptr + 3 * workload_idx;
55  const float *q_prime = Tj_qs_ptr + 3 * workload_idx;
56  const float *normal_p_prime =
57  Ri_normal_ps_ptr + 3 * workload_idx;
58 
59  float r = (p_prime[0] - q_prime[0]) * normal_p_prime[0] +
60  (p_prime[1] - q_prime[1]) * normal_p_prime[1] +
61  (p_prime[2] - q_prime[2]) * normal_p_prime[2];
62  if (abs(r) > threshold) return;
63 
64  float J_ij[12];
65  J_ij[0] = -q_prime[2] * normal_p_prime[1] +
66  q_prime[1] * normal_p_prime[2];
67  J_ij[1] = q_prime[2] * normal_p_prime[0] -
68  q_prime[0] * normal_p_prime[2];
69  J_ij[2] = -q_prime[1] * normal_p_prime[0] +
70  q_prime[0] * normal_p_prime[1];
71  J_ij[3] = normal_p_prime[0];
72  J_ij[4] = normal_p_prime[1];
73  J_ij[5] = normal_p_prime[2];
74  for (int k = 0; k < 6; ++k) {
75  J_ij[k + 6] = -J_ij[k];
76  }
77 
78  // Not optimized; Switch to reduction if necessary.
79 #if defined(BUILD_CUDA_MODULE) && defined(__CUDACC__)
80  for (int i_local = 0; i_local < 12; ++i_local) {
81  for (int j_local = 0; j_local < 12; ++j_local) {
82  atomicAdd(&AtA_local_ptr[i_local * 12 + j_local],
83  J_ij[i_local] * J_ij[j_local]);
84  }
85  atomicAdd(&Atb_local_ptr[i_local], J_ij[i_local] * r);
86  }
87  atomicAdd(residual_ptr, r * r);
88 #else
89 #pragma omp critical(FillInRigidAlignmentTermCPU)
90  {
91  for (int i_local = 0; i_local < 12; ++i_local) {
92  for (int j_local = 0; j_local < 12; ++j_local) {
93  AtA_local_ptr[i_local * 12 + j_local]
94  += J_ij[i_local] * J_ij[j_local];
95  }
96  Atb_local_ptr[i_local] += J_ij[i_local] * r;
97  }
98  *residual_ptr += r * r;
99  }
100 #endif
101  });
102 
103  // Then fill-in the large linear system
104  std::vector<int64_t> indices_vec(12);
105  for (int k = 0; k < 6; ++k) {
106  indices_vec[k] = i * 6 + k;
107  indices_vec[k + 6] = j * 6 + k;
108  }
109 
110  std::vector<int64_t> indices_i_vec;
111  std::vector<int64_t> indices_j_vec;
112  for (int local_i = 0; local_i < 12; ++local_i) {
113  for (int local_j = 0; local_j < 12; ++local_j) {
114  indices_i_vec.push_back(indices_vec[local_i]);
115  indices_j_vec.push_back(indices_vec[local_j]);
116  }
117  }
118 
119  core::Tensor indices(indices_vec, {12}, core::Int64, device);
120  core::Tensor indices_i(indices_i_vec, {12 * 12}, core::Int64, device);
121  core::Tensor indices_j(indices_j_vec, {12 * 12}, core::Int64, device);
122 
123  core::Tensor AtA_sub = AtA.IndexGet({indices_i, indices_j});
124  AtA.IndexSet({indices_i, indices_j}, AtA_sub + AtA_local.View({12 * 12}));
125 
126  core::Tensor Atb_sub = Atb.IndexGet({indices});
127  Atb.IndexSet({indices}, Atb_sub + Atb_local.View({12, 1}));
128 }
129 
130 #if defined(__CUDACC__)
131 void FillInSLACAlignmentTermCUDA
132 #else
134 #endif
135  (core::Tensor &AtA,
136  core::Tensor &Atb,
137  core::Tensor &residual,
138  const core::Tensor &Ti_Cps,
139  const core::Tensor &Tj_Cqs,
140  const core::Tensor &Cnormal_ps,
141  const core::Tensor &Ri_Cnormal_ps,
142  const core::Tensor &RjT_Ri_Cnormal_ps,
143  const core::Tensor &cgrid_idx_ps,
144  const core::Tensor &cgrid_idx_qs,
145  const core::Tensor &cgrid_ratio_qs,
146  const core::Tensor &cgrid_ratio_ps,
147  int i,
148  int j,
149  int n_frags,
150  float threshold) {
151  int64_t n = Ti_Cps.GetLength();
152  if (Tj_Cqs.GetLength() != n || Cnormal_ps.GetLength() != n ||
153  Ri_Cnormal_ps.GetLength() != n || RjT_Ri_Cnormal_ps.GetLength() != n ||
154  cgrid_idx_ps.GetLength() != n || cgrid_ratio_ps.GetLength() != n ||
155  cgrid_idx_qs.GetLength() != n || cgrid_ratio_qs.GetLength() != n) {
157  "Unable to setup linear system: input length mismatch.");
158  }
159 
160  int n_vars = Atb.GetLength();
161  float *AtA_ptr = static_cast<float *>(AtA.GetDataPtr());
162  float *Atb_ptr = static_cast<float *>(Atb.GetDataPtr());
163  float *residual_ptr = static_cast<float *>(residual.GetDataPtr());
164 
165  // Geometric properties
166  const float *Ti_Cps_ptr = static_cast<const float *>(Ti_Cps.GetDataPtr());
167  const float *Tj_Cqs_ptr = static_cast<const float *>(Tj_Cqs.GetDataPtr());
168  const float *Cnormal_ps_ptr =
169  static_cast<const float *>(Cnormal_ps.GetDataPtr());
170  const float *Ri_Cnormal_ps_ptr =
171  static_cast<const float *>(Ri_Cnormal_ps.GetDataPtr());
172  const float *RjT_Ri_Cnormal_ps_ptr =
173  static_cast<const float *>(RjT_Ri_Cnormal_ps.GetDataPtr());
174 
175  // Association properties
176  const int *cgrid_idx_ps_ptr =
177  static_cast<const int *>(cgrid_idx_ps.GetDataPtr());
178  const int *cgrid_idx_qs_ptr =
179  static_cast<const int *>(cgrid_idx_qs.GetDataPtr());
180  const float *cgrid_ratio_ps_ptr =
181  static_cast<const float *>(cgrid_ratio_ps.GetDataPtr());
182  const float *cgrid_ratio_qs_ptr =
183  static_cast<const float *>(cgrid_ratio_qs.GetDataPtr());
184 
186  AtA.GetDevice(), n, [=] OPEN3D_DEVICE(int64_t workload_idx) {
187  const float *Ti_Cp = Ti_Cps_ptr + 3 * workload_idx;
188  const float *Tj_Cq = Tj_Cqs_ptr + 3 * workload_idx;
189  const float *Cnormal_p = Cnormal_ps_ptr + 3 * workload_idx;
190  const float *Ri_Cnormal_p =
191  Ri_Cnormal_ps_ptr + 3 * workload_idx;
192  const float *RjTRi_Cnormal_p =
193  RjT_Ri_Cnormal_ps_ptr + 3 * workload_idx;
194 
195  const int *cgrid_idx_p = cgrid_idx_ps_ptr + 8 * workload_idx;
196  const int *cgrid_idx_q = cgrid_idx_qs_ptr + 8 * workload_idx;
197  const float *cgrid_ratio_p =
198  cgrid_ratio_ps_ptr + 8 * workload_idx;
199  const float *cgrid_ratio_q =
200  cgrid_ratio_qs_ptr + 8 * workload_idx;
201 
202  float r = (Ti_Cp[0] - Tj_Cq[0]) * Ri_Cnormal_p[0] +
203  (Ti_Cp[1] - Tj_Cq[1]) * Ri_Cnormal_p[1] +
204  (Ti_Cp[2] - Tj_Cq[2]) * Ri_Cnormal_p[2];
205  if (abs(r) > threshold) return;
206 
207  // Now we fill in a 60 x 60 sub-matrix: 2 x (6 + 8 x 3)
208  float J[60];
209  int idx[60];
210 
211  // Jacobian w.r.t. Ti: 0-6
212  J[0] = -Tj_Cq[2] * Ri_Cnormal_p[1] + Tj_Cq[1] * Ri_Cnormal_p[2];
213  J[1] = Tj_Cq[2] * Ri_Cnormal_p[0] - Tj_Cq[0] * Ri_Cnormal_p[2];
214  J[2] = -Tj_Cq[1] * Ri_Cnormal_p[0] + Tj_Cq[0] * Ri_Cnormal_p[1];
215  J[3] = Ri_Cnormal_p[0];
216  J[4] = Ri_Cnormal_p[1];
217  J[5] = Ri_Cnormal_p[2];
218 
219  // Jacobian w.r.t. Tj: 6-12
220  for (int k = 0; k < 6; ++k) {
221  J[k + 6] = -J[k];
222 
223  idx[k + 0] = 6 * i + k;
224  idx[k + 6] = 6 * j + k;
225  }
226 
227  // Jacobian w.r.t. C over p: 12-36
228  for (int k = 0; k < 8; ++k) {
229  J[12 + k * 3 + 0] = cgrid_ratio_p[k] * Cnormal_p[0];
230  J[12 + k * 3 + 1] = cgrid_ratio_p[k] * Cnormal_p[1];
231  J[12 + k * 3 + 2] = cgrid_ratio_p[k] * Cnormal_p[2];
232 
233  idx[12 + k * 3 + 0] = 6 * n_frags + cgrid_idx_p[k] * 3 + 0;
234  idx[12 + k * 3 + 1] = 6 * n_frags + cgrid_idx_p[k] * 3 + 1;
235  idx[12 + k * 3 + 2] = 6 * n_frags + cgrid_idx_p[k] * 3 + 2;
236  }
237 
238  // Jacobian w.r.t. C over q: 36-60
239  for (int k = 0; k < 8; ++k) {
240  J[36 + k * 3 + 0] = -cgrid_ratio_q[k] * RjTRi_Cnormal_p[0];
241  J[36 + k * 3 + 1] = -cgrid_ratio_q[k] * RjTRi_Cnormal_p[1];
242  J[36 + k * 3 + 2] = -cgrid_ratio_q[k] * RjTRi_Cnormal_p[2];
243 
244  idx[36 + k * 3 + 0] = 6 * n_frags + cgrid_idx_q[k] * 3 + 0;
245  idx[36 + k * 3 + 1] = 6 * n_frags + cgrid_idx_q[k] * 3 + 1;
246  idx[36 + k * 3 + 2] = 6 * n_frags + cgrid_idx_q[k] * 3 + 2;
247  }
248 
249  // Not optimized; Switch to reduction if necessary.
250 #if defined(__CUDACC__)
251  for (int ki = 0; ki < 60; ++ki) {
252  for (int kj = 0; kj < 60; ++kj) {
253  float AtA_ij = J[ki] * J[kj];
254  int ij = idx[ki] * n_vars + idx[kj];
255  atomicAdd(AtA_ptr + ij, AtA_ij);
256  }
257  float Atb_i = J[ki] * r;
258  atomicAdd(Atb_ptr + idx[ki], Atb_i);
259  }
260  atomicAdd(residual_ptr, r * r);
261 #else
262 #pragma omp critical(FillInSLACAlignmentTermCPU)
263  {
264  for (int ki = 0; ki < 60; ++ki) {
265  for (int kj = 0; kj < 60; ++kj) {
266  AtA_ptr[idx[ki] * n_vars + idx[kj]]
267  += J[ki] * J[kj];
268  }
269  Atb_ptr[idx[ki]] += J[ki] * r;
270  }
271  *residual_ptr += r * r;
272  }
273 #endif
274  });
275 }
276 
277 #if defined(__CUDACC__)
278 void FillInSLACRegularizerTermCUDA
279 #else
281 #endif
282  (core::Tensor &AtA,
283  core::Tensor &Atb,
284  core::Tensor &residual,
285  const core::Tensor &grid_idx,
286  const core::Tensor &grid_nbs_idx,
287  const core::Tensor &grid_nbs_mask,
288  const core::Tensor &positions_init,
289  const core::Tensor &positions_curr,
290  float weight,
291  int n_frags,
292  int anchor_idx) {
293 
294  int64_t n = grid_idx.GetLength();
295  int64_t n_vars = Atb.GetLength();
296 
297  float *AtA_ptr = static_cast<float *>(AtA.GetDataPtr());
298  float *Atb_ptr = static_cast<float *>(Atb.GetDataPtr());
299  float *residual_ptr = static_cast<float *>(residual.GetDataPtr());
300 
301  const int *grid_idx_ptr = static_cast<const int *>(grid_idx.GetDataPtr());
302  const int *grid_nbs_idx_ptr =
303  static_cast<const int *>(grid_nbs_idx.GetDataPtr());
304  const bool *grid_nbs_mask_ptr =
305  static_cast<const bool *>(grid_nbs_mask.GetDataPtr());
306 
307  const float *positions_init_ptr =
308  static_cast<const float *>(positions_init.GetDataPtr());
309  const float *positions_curr_ptr =
310  static_cast<const float *>(positions_curr.GetDataPtr());
311 
313  AtA.GetDevice(), n, [=] OPEN3D_DEVICE(int64_t workload_idx) {
314  // Enumerate 6 neighbors
315  int idx_i = grid_idx_ptr[workload_idx];
316 
317  const int *idx_nbs = grid_nbs_idx_ptr + 6 * workload_idx;
318  const bool *mask_nbs = grid_nbs_mask_ptr + 6 * workload_idx;
319 
320  // Build a 3x3 linear system to compute the local R
321  float cov[3][3] = {{0}};
322  float U[3][3], V[3][3], S[3];
323 
324  int cnt = 0;
325  for (int k = 0; k < 6; ++k) {
326  bool mask_k = mask_nbs[k];
327  if (!mask_k) continue;
328 
329  int idx_k = idx_nbs[k];
330 
331  // Now build linear systems
332  float diff_ik_init[3] = {
333  positions_init_ptr[idx_i * 3 + 0] -
334  positions_init_ptr[idx_k * 3 + 0],
335  positions_init_ptr[idx_i * 3 + 1] -
336  positions_init_ptr[idx_k * 3 + 1],
337  positions_init_ptr[idx_i * 3 + 2] -
338  positions_init_ptr[idx_k * 3 + 2]};
339  float diff_ik_curr[3] = {
340  positions_curr_ptr[idx_i * 3 + 0] -
341  positions_curr_ptr[idx_k * 3 + 0],
342  positions_curr_ptr[idx_i * 3 + 1] -
343  positions_curr_ptr[idx_k * 3 + 1],
344  positions_curr_ptr[idx_i * 3 + 2] -
345  positions_curr_ptr[idx_k * 3 + 2]};
346 
347  // Build linear system by computing XY^T when formulating Y
348  // = RX Y: curr X: init
349  for (int i = 0; i < 3; ++i) {
350  for (int j = 0; j < 3; ++j) {
351  cov[i][j] += diff_ik_init[i] * diff_ik_curr[j];
352  }
353  }
354  ++cnt;
355  }
356 
357  if (cnt < 3) {
358  return;
359  }
360 
361  core::linalg::kernel::svd3x3(*cov, *U, S, *V);
362 
363  float R[3][3];
366 
367  float d = core::linalg::kernel::det3x3(*R);
368 
369  if (d < 0) {
370  U[2][0] = -U[2][0];
371  U[2][1] = -U[2][1];
372  U[2][2] = -U[2][2];
374  }
375 
376  // Now we have R, we build Hessian and residuals
377  // But first, we need to anchor a point
378  if (idx_i == anchor_idx) {
379  R[0][0] = R[1][1] = R[2][2] = 1;
380  R[0][1] = R[0][2] = R[1][0] = R[1][2] = R[2][0] = R[2][1] =
381  0;
382  }
383  for (int k = 0; k < 6; ++k) {
384  bool mask_k = mask_nbs[k];
385 
386  if (mask_k) {
387  int idx_k = idx_nbs[k];
388 
389  float diff_ik_init[3] = {
390  positions_init_ptr[idx_i * 3 + 0] -
391  positions_init_ptr[idx_k * 3 + 0],
392  positions_init_ptr[idx_i * 3 + 1] -
393  positions_init_ptr[idx_k * 3 + 1],
394  positions_init_ptr[idx_i * 3 + 2] -
395  positions_init_ptr[idx_k * 3 + 2]};
396  float diff_ik_curr[3] = {
397  positions_curr_ptr[idx_i * 3 + 0] -
398  positions_curr_ptr[idx_k * 3 + 0],
399  positions_curr_ptr[idx_i * 3 + 1] -
400  positions_curr_ptr[idx_k * 3 + 1],
401  positions_curr_ptr[idx_i * 3 + 2] -
402  positions_curr_ptr[idx_k * 3 + 2]};
403  float R_diff_ik_curr[3];
404 
405  core::linalg::kernel::matmul3x3_3x1(*R, diff_ik_init,
406  R_diff_ik_curr);
407 
408  float local_r[3];
409  local_r[0] = diff_ik_curr[0] - R_diff_ik_curr[0];
410  local_r[1] = diff_ik_curr[1] - R_diff_ik_curr[1];
411  local_r[2] = diff_ik_curr[2] - R_diff_ik_curr[2];
412 
413  int offset_idx_i = 3 * idx_i + 6 * n_frags;
414  int offset_idx_k = 3 * idx_k + 6 * n_frags;
415 
416 #if defined(__CUDACC__)
417  // Update residual
418  atomicAdd(residual_ptr,
419  weight * (local_r[0] * local_r[0] +
420  local_r[1] * local_r[1] +
421  local_r[2] * local_r[2]));
422 
423  for (int axis = 0; axis < 3; ++axis) {
424  // Update AtA: 2x2
425  atomicAdd(&AtA_ptr[(offset_idx_i + axis) * n_vars +
426  offset_idx_i + axis],
427  weight);
428  atomicAdd(&AtA_ptr[(offset_idx_k + axis) * n_vars +
429  offset_idx_k + axis],
430  weight);
431  atomicAdd(&AtA_ptr[(offset_idx_i + axis) * n_vars +
432  offset_idx_k + axis],
433  -weight);
434  atomicAdd(&AtA_ptr[(offset_idx_k + axis) * n_vars +
435  offset_idx_i + axis],
436  -weight);
437 
438  // Update Atb: 2x1
439  atomicAdd(&Atb_ptr[offset_idx_i + axis],
440  +weight * local_r[axis]);
441  atomicAdd(&Atb_ptr[offset_idx_k + axis],
442  -weight * local_r[axis]);
443  }
444 #else
445 #pragma omp critical(FillInSLACRegularizerTermCPU)
446  {
447  // Update residual
448  *residual_ptr += weight * (local_r[0] * local_r[0] +
449  local_r[1] * local_r[1] +
450  local_r[2] * local_r[2]);
451 
452  for (int axis = 0; axis < 3; ++axis) {
453  // Update AtA: 2x2
454  AtA_ptr[(offset_idx_i + axis) * n_vars +
455  offset_idx_i + axis] += weight;
456  AtA_ptr[(offset_idx_k + axis) * n_vars +
457  offset_idx_k + axis] += weight;
458 
459  AtA_ptr[(offset_idx_i + axis) * n_vars +
460  offset_idx_k + axis] -= weight;
461  AtA_ptr[(offset_idx_k + axis) * n_vars +
462  offset_idx_i + axis] -= weight;
463 
464  // Update Atb: 2x1
465  Atb_ptr[offset_idx_i + axis] += weight * local_r[axis];
466  Atb_ptr[offset_idx_k + axis] -= weight * local_r[axis];
467  }
468  }
469 #endif
470  }
471  }
472  });
473 }
474 } // namespace kernel
475 } // namespace pipelines
476 } // namespace t
477 } // namespace open3d
#define OPEN3D_DEVICE
Definition: CUDAUtils.h:45
#define LogError(...)
Definition: Logging.h:48
Definition: Device.h:18
Definition: Tensor.h:32
T * GetDataPtr()
Definition: Tensor.h:1133
Tensor View(const SizeVector &dst_shape) const
Definition: Tensor.cpp:688
static Tensor Zeros(const SizeVector &shape, Dtype dtype, const Device &device=Device("CPU:0"))
Create a tensor fill with zeros.
Definition: Tensor.cpp:373
void IndexSet(const std::vector< Tensor > &index_tensors, const Tensor &src_tensor)
Advanced indexing getter.
Definition: Tensor.cpp:903
Tensor IndexGet(const std::vector< Tensor > &index_tensors) const
Advanced indexing getter. This will always allocate a new Tensor.
Definition: Tensor.cpp:872
OPEN3D_DEVICE OPEN3D_FORCE_INLINE void transpose3x3_(scalar_t *A_3x3)
Definition: Matrix.h:151
OPEN3D_DEVICE OPEN3D_FORCE_INLINE void matmul3x3_3x3(const scalar_t *A_3x3, const scalar_t *B_3x3, scalar_t *C_3x3)
Definition: Matrix.h:48
OPEN3D_DEVICE OPEN3D_FORCE_INLINE void svd3x3(const scalar_t *A_3x3, scalar_t *U_3x3, scalar_t *S_3x1, scalar_t *V_3x3)
OPEN3D_DEVICE OPEN3D_FORCE_INLINE scalar_t det3x3(const scalar_t *A_3x3)
Definition: Matrix.h:89
const Dtype Int64
Definition: Dtype.cpp:47
void ParallelFor(const Device &device, int64_t n, const func_t &func)
Definition: ParallelFor.h:103
const Dtype Float32
Definition: Dtype.cpp:42
void FillInSLACAlignmentTermCPU(core::Tensor &AtA, core::Tensor &Atb, core::Tensor &residual, const core::Tensor &Ti_qs, const core::Tensor &Tj_qs, const core::Tensor &normal_ps, const core::Tensor &Ri_normal_ps, const core::Tensor &RjT_Ri_normal_ps, const core::Tensor &cgrid_idx_ps, const core::Tensor &cgrid_idx_qs, const core::Tensor &cgrid_ratio_qs, const core::Tensor &cgrid_ratio_ps, int i, int j, int n, float threshold)
Definition: FillInLinearSystemImpl.h:135
void FillInRigidAlignmentTermCPU(core::Tensor &AtA, core::Tensor &Atb, core::Tensor &residual, const core::Tensor &Ti_qs, const core::Tensor &Tj_qs, const core::Tensor &Ri_normal_ps, int i, int j, float threshold)
Definition: FillInLinearSystemImpl.h:21
void FillInSLACRegularizerTermCPU(core::Tensor &AtA, core::Tensor &Atb, core::Tensor &residual, const core::Tensor &grid_idx, const core::Tensor &grid_nbs_idx, const core::Tensor &grid_nbs_mask, const core::Tensor &positions_init, const core::Tensor &positions_curr, float weight, int n, int anchor_idx)
Definition: FillInLinearSystemImpl.h:282
Definition: PinholeCameraIntrinsic.cpp:16