text stringlengths 0 1.47k |
|---|
# means some twist leaves the overlap unchanged → rotation genuinely unconstrained. Kept as a low |
# gate so it fires only on extreme rotational flatness (the residual gate catches ordinary |
# failures); a discretely-sampled near-symmetric remnant that STILL aligns (Chamfer-success) is |
# intentionally NOT failed by this gate. |
_FS_AMBIG_SENSITIVITY = 0.02 |
# Diagnostic only: RANSAC hypothesis rotation spread among near-best hypotheses. |
_FS_AMBIG_ROT_SPREAD_DEG = 12.0 # near-best hypotheses spread above this (deg) → rotation disagreement |
_FS_TOPK_HYP = 12 # number of top RANSAC hypotheses kept for the spread test |
_FS_HYP_SCORE_SLACK = 0.85 # a hypothesis is "near-best" if inliers >= slack * best_inliers |
# Score epsilon for stability tie-break in global_align (used in align.py) |
_SCORE_EPS = 1e-3 |
# ── k-NN ────────────────────────────────────────────────────────────────────── |
def _knn_indices(pts: torch.Tensor, k: int) -> torch.Tensor: |
"""For each point return the indices of its ``k`` nearest neighbours (excluding self). |
Uses chunked ``cdist`` to bound memory. Returns ``(N, k)`` int64 indices. |
CPU-friendly (no cuda required). |
""" |
N = pts.shape[0] |
k = min(k, N - 1) |
out = pts.new_empty((N, k), dtype=torch.int64) |
step = max(1, int(4_000_000 // max(N, 1))) |
for lo in range(0, N, step): |
hi = min(lo + step, N) |
d = torch.cdist(pts[lo:hi], pts) # (chunk, N) |
d[:, lo:hi].fill_diagonal_(float("inf")) # exclude self |
_, idx = torch.topk(d, k, dim=1, largest=False) |
out[lo:hi] = idx |
return out |
# ── normals ──────────────────────────────────────────────────────────────────── |
def _estimate_normals(pts: torch.Tensor, idx: torch.Tensor) -> torch.Tensor: |
"""Per-point surface normal ``(N, 3)`` from the smallest-eigenvector of the local covariance. |
The normal is the eigenvector of the k-NN covariance with the *smallest* eigenvalue (the |
direction of least variance = surface normal). Signs are disambiguated to point away from the |
cloud centroid so both source and target orient consistently (FPFH features use the normal |
directions, so a consistent convention is required for the descriptors to match across clouds). |
Args: |
pts: ``(N, 3)`` float32 cloud. |
idx: ``(N, k)`` precomputed k-NN indices. |
Returns: |
``(N, 3)`` unit normals on the same device. |
""" |
N = pts.shape[0] |
nbrs = pts[idx] # (N, k, 3) |
centred = nbrs - nbrs.mean(dim=1, keepdim=True) # (N, k, 3) |
cov = torch.einsum("nki,nkj->nij", centred, centred) / centred.shape[1] # (N, 3, 3) |
# eigh returns ascending eigenvalues; smallest -> column 0 of eigenvectors |
try: |
_, evecs = torch.linalg.eigh(cov) # evecs: (N, 3, 3), columns are eigenvectors |
except Exception: |
return torch.zeros(N, 3, device=pts.device, dtype=pts.dtype).index_fill_( |
1, torch.tensor([2], device=pts.device), 1.0 |
) |
normals = evecs[:, :, 0] # (N, 3) smallest-eigenvalue direction |
# Orient away from the cloud centroid (consistent sign convention across clouds). |
centroid = pts.mean(dim=0, keepdim=True) # (1, 3) |
out = pts - centroid # (N, 3) outward radial direction |
flip = (normals * out).sum(dim=1) < 0.0 # normal points inward → flip |
normals = torch.where(flip.unsqueeze(1), -normals, normals) |
normals = normals / normals.norm(dim=1, keepdim=True).clamp_min(1e-9) |
return normals |
# ── FPFH descriptor (Rusu et al., ICRA 2009) ─────────────────────────────────── |
def _spfh(pts: torch.Tensor, normals: torch.Tensor, idx: torch.Tensor, bins: int) -> torch.Tensor: |
"""Simplified Point Feature Histogram ``(N, 3*bins)`` from the Darboux-frame angles. |
For every (anchor p, neighbour q) pair we build the Darboux frame from the two estimated |
normals and compute the three classic PFH features: |
u = n_p |
v = (q - p) / ||q - p|| × u (cross product, then normalised) |
w = u × v |
alpha = v · n_q (in [-1, 1]) |
phi = u · (q - p) / ||q - p|| (in [-1, 1]) |
theta = atan2(w · n_q, u · n_q) (in [-pi, pi]) |
Each feature is histogrammed into ``bins`` bins; the three histograms are concatenated and |
L1-normalised. Translation/rotation invariant by construction (only relative directions enter). |
Args: |
pts: ``(N, 3)`` cloud. |
normals: ``(N, 3)`` unit normals. |
idx: ``(N, k)`` k-NN indices. |
bins: histogram bins per feature. |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.