STEM / code /3dSAGER /alignment.py
eduzrh
feat: STEM benchmark initial release
bcb16da
Raw
History Blame Contribute Delete
20.5 kB
"""
alignment.py
Two-stage rigid 3D alignment of candidate buildings onto the index coordinate frame.
Stage 1 — Estimate transform (Arun et al. 1987):
Selects high-confidence matched pairs as anchors, estimates rotation R and
translation t via SVD least-squares, then validates the result by computing
per-anchor residuals. If the mean residual exceeds the configured threshold
the alignment is rejected and the pipeline continues with geometric scores only.
Stage 2 — Re-score and output:
Applies (R, t) to all cand centroids and re-scores every pair as:
final_score = alpha * geometric_score + (1 - alpha) * spatial_score
where spatial_score = 1 / (1 + distance_after_alignment).
Then applies (R, t) to all cand geometry and writes an aligned CityJSON file.
Reference:
K. S. Arun, T. S. Huang, and S. D. Blostein. 1987.
Least-Squares Fitting of Two 3-D Point Sets.
IEEE TPAMI 9(5):698-700. doi:10.1109/TPAMI.1987.4767965
Usage:
from alignment import RigidAligner
import config
aligner = RigidAligner(config.Alignment, logger=logger)
rescored_pairs = aligner.run(
object_dict,
scored_pairs, # list of (cand_id, index_id, geometric_score)
suffix="seed1",
ground_truth_R=simulator.R_crs, # optional, for evaluation only
ground_truth_t=simulator.t_crs,
)
# rescored_pairs: list of (cand_id, index_id, final_score)
# results/aligned_candidates_seed1.json written if alignment succeeded
"""
import json
import os
import logging
import numpy as np
from typing import List, Tuple, Optional
import config as cfg
class RigidAligner:
"""
Estimates a rigid 3D transform from anchor pairs and re-scores all matches.
Parameters
----------
align_config : config.Alignment (class reference)
logger : logging.Logger (optional)
"""
def __init__(self, align_config=None, logger=None):
if align_config is None:
align_config = cfg.Alignment
self.enabled = align_config.enabled
self.min_anchor_pairs = align_config.min_anchor_pairs
self.confidence_threshold = align_config.confidence_threshold
self.max_residual_threshold = align_config.max_residual_threshold
self.alpha = align_config.alpha
self.output_crs = align_config.output_crs
self.use_ransac = getattr(align_config, 'use_ransac', True)
self.ransac_iterations = getattr(align_config, 'ransac_iterations', 1000)
self.ransac_inlier_threshold = getattr(align_config, 'ransac_inlier_threshold', 10.0)
self.spatial_sigma = getattr(align_config, 'spatial_sigma', 3.0)
self.logger = logger or logging.getLogger(__name__)
# Set after a successful alignment
self.R: Optional[np.ndarray] = None
self.t: Optional[np.ndarray] = None
self.mean_residual: Optional[float] = None
self.n_anchors: int = 0
self.alignment_succeeded: bool = False
# ------------------------------------------------------------------ #
# Public API
# ------------------------------------------------------------------ #
def run(
self,
object_dict: dict,
scored_pairs: List[Tuple[str, str, float]],
suffix: str = "",
ground_truth_R: Optional[np.ndarray] = None,
ground_truth_t: Optional[np.ndarray] = None,
) -> List[Tuple[str, str, float]]:
"""
Full alignment pipeline: estimate transform → validate → re-score → output.
Parameters
----------
object_dict : dict
Full object dict with 'cands' and 'index'.
scored_pairs : list of (cand_id, index_id, geometric_score)
All test pairs with their classifier probability scores.
suffix : str
Appended to the output filename.
ground_truth_R, ground_truth_t : np.ndarray, optional
If provided (from DisasterSimulator), the alignment error is logged
for evaluation purposes. Not used in the alignment itself.
Returns
-------
list of (cand_id, index_id, final_score)
If alignment succeeded: final_score = alpha*geometric + (1-alpha)*spatial.
If alignment failed/skipped: final_score = geometric_score unchanged.
"""
if not self.enabled:
self.logger.info("[RigidAligner] Disabled — returning geometric scores.")
return scored_pairs
# --- Stage 1: Estimate transform ---
anchors = self._select_anchors(scored_pairs, object_dict)
self.n_anchors = len(anchors)
if self.n_anchors < self.min_anchor_pairs:
self.logger.warning(
f"[RigidAligner] Only {self.n_anchors} anchor pairs found "
f"(need {self.min_anchor_pairs}, threshold={self.confidence_threshold}). "
f"Skipping alignment — returning geometric scores."
)
return scored_pairs
P, Q = self._build_point_sets(anchors, object_dict)
if self.use_ransac:
R, t, n_inliers, inlier_mask = self._estimate_rigid_transform_ransac(P, Q)
# Validate using INLIER mean residual when RANSAC found enough inliers.
# The all-anchor mean is dominated by false-positive anchors (look-alike
# buildings the classifier scored high) and unfairly rejects a correct
# transform whenever anchor-pool precision is low.
if n_inliers >= self.min_anchor_pairs and inlier_mask.any():
self.mean_residual = self._compute_residual(P[inlier_mask], Q[inlier_mask], R, t)
all_anchor_mean = self._compute_residual(P, Q, R, t)
self.logger.info(
f"[RigidAligner] RANSAC: {self.n_anchors} anchors | "
f"{n_inliers} inliers | inlier mean residual = {self.mean_residual:.2f} m "
f"(all-anchor mean = {all_anchor_mean:.2f} m)"
)
else:
self.mean_residual = self._compute_residual(P, Q, R, t)
self.logger.info(
f"[RigidAligner] RANSAC: {self.n_anchors} anchors | "
f"only {n_inliers} inliers (< {self.min_anchor_pairs}) | "
f"all-anchor mean residual = {self.mean_residual:.2f} m"
)
else:
R, t = self._estimate_rigid_transform(P, Q)
self.mean_residual = self._compute_residual(P, Q, R, t)
self.logger.info(
f"[RigidAligner] {self.n_anchors} anchors | "
f"mean residual = {self.mean_residual:.2f} m"
)
if self.mean_residual > self.max_residual_threshold:
self.logger.warning(
f"[RigidAligner] Mean residual {self.mean_residual:.2f} m exceeds "
f"threshold {self.max_residual_threshold} m. "
f"Alignment rejected — returning geometric scores."
)
return scored_pairs
self.R, self.t = R, t
self.alignment_succeeded = True
# Optional: log error vs ground-truth transform
if ground_truth_R is not None and ground_truth_t is not None:
self._log_ground_truth_error(ground_truth_R, ground_truth_t)
# --- Stage 2: Re-score using aligned positions ---
rescored = self._rescore_pairs(scored_pairs, object_dict)
# Apply transform to full cand geometry and write output
self._apply_transform_to_geometry(object_dict['cands'])
self._write_cityjson(object_dict['cands'], suffix)
return rescored
# ------------------------------------------------------------------ #
# Stage 1 helpers
# ------------------------------------------------------------------ #
def _select_anchors(
self,
scored_pairs: List[Tuple[str, str, float]],
object_dict: dict,
) -> List[Tuple[str, str]]:
"""Return (cand_id, index_id) pairs with score >= confidence_threshold."""
cands_keys = set(object_dict['cands'].keys())
index_keys = set(object_dict['index'].keys())
return [
(cid, iid)
for cid, iid, score in scored_pairs
if score >= self.confidence_threshold
and cid in cands_keys
and iid in index_keys
]
@staticmethod
def _build_point_sets(
anchors: List[Tuple[str, str]],
object_dict: dict,
) -> Tuple[np.ndarray, np.ndarray]:
"""
Build point sets from anchor centroids.
P = index centroids (target), Q = cand centroids (source).
Goal: find R, t such that P_i ≈ R @ Q_i + t.
"""
P = np.array([object_dict['index'][iid]['centroid'] for _, iid in anchors], dtype=np.float64)
Q = np.array([object_dict['cands'][cid]['centroid'] for cid, _ in anchors], dtype=np.float64)
return P, Q
@staticmethod
def _estimate_rigid_transform(
P: np.ndarray,
Q: np.ndarray,
) -> Tuple[np.ndarray, np.ndarray]:
"""
Arun et al. 1987 SVD least-squares rigid transform.
Returns R (3x3) and t (3,) such that P_i ≈ R @ Q_i + t.
"""
p_bar = P.mean(axis=0) # (3,)
q_bar = Q.mean(axis=0) # (3,)
P_prime = P - p_bar # (N, 3) centered
Q_prime = Q - q_bar # (N, 3) centered
H = Q_prime.T @ P_prime # (3, 3) cross-covariance
U, _, Vt = np.linalg.svd(H)
V = Vt.T
R = V @ U.T
# Fix reflection (degenerate / coplanar case)
if np.linalg.det(R) < 0:
V[:, 2] *= -1
R = V @ U.T
t = p_bar - R @ q_bar
return R, t
def _estimate_rigid_transform_ransac(
self,
P: np.ndarray,
Q: np.ndarray,
):
"""
RANSAC-based rigid transform estimation.
Each iteration samples 3 anchor pairs, estimates R and t via SVD,
then counts how many of all anchors are consistent (inliers) under
that transform. The best transform (most inliers) is refitted on
all its inliers via SVD for a final least-squares solution.
Parameters
----------
P : (N, 3) index centroids
Q : (N, 3) cand centroids
Returns
-------
R : (3, 3) rotation matrix
t : (3,) translation vector
n_inliers : int
"""
n = len(P)
best_inlier_mask = np.zeros(n, dtype=bool)
best_n_inliers = 0
best_R, best_t = self._estimate_rigid_transform(P, Q) # fallback
rng = np.random.default_rng(42)
for _ in range(self.ransac_iterations):
# Sample 3 unique anchor pairs
idx = rng.choice(n, size=3, replace=False)
P_sample, Q_sample = P[idx], Q[idx]
# Skip degenerate (collinear) samples
if np.linalg.matrix_rank(P_sample - P_sample.mean(axis=0)) < 2:
continue
R_cand, t_cand = self._estimate_rigid_transform(P_sample, Q_sample)
# Count inliers: anchors whose residual < threshold under this transform
P_hat = (R_cand @ Q.T).T + t_cand
residuals = np.linalg.norm(P_hat - P, axis=1)
inlier_mask = residuals < self.ransac_inlier_threshold
n_inliers = inlier_mask.sum()
if n_inliers > best_n_inliers:
best_n_inliers = n_inliers
best_inlier_mask = inlier_mask
best_R, best_t = R_cand, t_cand
# Refit on all inliers of the best solution
if best_n_inliers >= 3:
best_R, best_t = self._estimate_rigid_transform(
P[best_inlier_mask], Q[best_inlier_mask]
)
# Recompute inlier mask under the refit transform so it reflects the
# final R, t rather than the 3-sample one used to score iterations.
P_hat = (best_R @ Q.T).T + best_t
residuals = np.linalg.norm(P_hat - P, axis=1)
best_inlier_mask = residuals < self.ransac_inlier_threshold
best_n_inliers = int(best_inlier_mask.sum())
self.logger.info(
f"[RigidAligner] RANSAC refit on {best_n_inliers}/{n} inliers "
f"(threshold={self.ransac_inlier_threshold} m, "
f"iterations={self.ransac_iterations})"
)
else:
self.logger.warning(
f"[RigidAligner] RANSAC found only {best_n_inliers} inliers — "
f"falling back to full SVD"
)
return best_R, best_t, best_n_inliers, best_inlier_mask
@staticmethod
def _compute_residual(
P: np.ndarray,
Q: np.ndarray,
R: np.ndarray,
t: np.ndarray,
) -> float:
"""Mean Euclidean residual ||R @ q_i + t - p_i|| over all anchor pairs."""
P_hat = (R @ Q.T).T + t # (N, 3)
residuals = np.linalg.norm(P_hat - P, axis=1)
return float(residuals.mean())
def _log_ground_truth_error(
self,
gt_R: np.ndarray,
gt_t: np.ndarray,
) -> None:
"""
Log rotation and translation error vs the ground-truth CRS transform.
The disaster simulator applies: cand = R_crs @ original + t_crs
So the aligner should recover the INVERSE transform:
self.R ≈ R_crs^T (so that R_crs^T @ R_crs = I)
self.t ≈ -R_crs^T @ t_crs
Rotation check: self.R @ gt_R should be close to I.
Translation check: self.t should be close to -gt_R^T @ gt_t.
"""
# Rotation error: R_recovered @ R_gt should equal I if perfect
R_check = self.R @ gt_R
angle_err = np.degrees(np.arccos(
np.clip((np.trace(R_check) - 1.0) / 2.0, -1.0, 1.0)
))
# Translation error: compare recovered t to the expected inverse translation
expected_t = -gt_R.T @ gt_t
t_err = np.linalg.norm(self.t - expected_t)
self.logger.info(
f"[RigidAligner] Ground-truth comparison: "
f"rotation error = {angle_err:.2f}°, "
f"translation error = {t_err:.1f} m"
)
# ------------------------------------------------------------------ #
# Stage 2 helpers
# ------------------------------------------------------------------ #
def _rescore_pairs(
self,
scored_pairs: List[Tuple[str, str, float]],
object_dict: dict,
) -> List[Tuple[str, str, float]]:
"""
Re-score pairs combining geometric score with spatial proximity
after aligning cand centroids to the index frame.
final_score = alpha * geometric_score + (1 - alpha) * spatial_score
spatial_score = exp(- d² / (2 · spatial_sigma²)) # Gaussian, σ default 3 m
"""
# Apply R, t to cand centroids only (fast; full geometry updated later)
aligned_cand_centroids = {
bid: self.R @ np.asarray(data['centroid'], dtype=np.float64) + self.t
for bid, data in object_dict['cands'].items()
}
index_centroids = {
bid: np.asarray(data['centroid'], dtype=np.float64)
for bid, data in object_dict['index'].items()
}
rescored = []
for cid, iid, geo_score in scored_pairs:
if cid in aligned_cand_centroids and iid in index_centroids:
dist = float(np.linalg.norm(aligned_cand_centroids[cid] - index_centroids[iid]))
# Gaussian decay with σ = spatial_sigma (default 3 m, ≈ median true-match
# residual). Sharply suppresses look-alikes that land 5–10 m away while
# giving high scores to genuine matches at d ≤ σ.
spatial_score = float(np.exp(-(dist * dist) / (2.0 * self.spatial_sigma ** 2)))
final_score = self.alpha * geo_score + (1.0 - self.alpha) * spatial_score
else:
final_score = geo_score # fallback if ID missing
rescored.append((cid, iid, final_score))
# Log score distribution summary
geo_scores = [s for _, _, s in scored_pairs]
final_scores = [s for _, _, s in rescored]
self.logger.info(
f"[RigidAligner] Score re-scaling: "
f"geometric mean={np.mean(geo_scores):.3f} → "
f"final mean={np.mean(final_scores):.3f} "
f"(alpha={self.alpha})"
)
return rescored
def _apply_transform_to_geometry(self, cands: dict) -> None:
"""Apply (R, t) to all cand vertices, centroids, and polygon_mesh."""
for building in cands.values():
verts = building['vertices']
building['vertices'] = (self.R @ verts.T).T + self.t
building['centroid'] = self.R @ np.asarray(building['centroid'], dtype=np.float64) + self.t
new_mesh = []
for surface in building['polygon_mesh']:
new_surface = [(self.R @ np.array(c, dtype=np.float64) + self.t).tolist() for c in surface]
new_mesh.append(new_surface)
building['polygon_mesh'] = new_mesh
# ------------------------------------------------------------------ #
# CityJSON output
# ------------------------------------------------------------------ #
def _write_cityjson(self, cands: dict, suffix: str) -> None:
"""
Write aligned candidates to a CityJSON 1.1 file in the index CRS.
Vertices are stored as floating-point world coordinates (no re-quantization).
Each building is written as a Solid LOD2 geometry preserving the original
surface structure.
"""
results_dir = cfg.FilePaths.results_path
os.makedirs(results_dir, exist_ok=True)
out_path = os.path.join(results_dir, f"aligned_candidates_{suffix}.json")
city_objects = {}
all_vertices = []
vertex_index = {} # (x, y, z) rounded → global index
epsg_code = self.output_crs.split(":")[-1]
for bid, building in cands.items():
verts = building['vertices'] # (N, 3) already aligned
# Build local vertex index
local_idx = {}
for v in verts:
key = (round(float(v[0]), 6), round(float(v[1]), 6), round(float(v[2]), 6))
if key not in vertex_index:
vertex_index[key] = len(all_vertices)
all_vertices.append(list(key))
local_idx[key] = vertex_index[key]
# Encode polygon_mesh as surface boundary index lists
boundaries = []
for surface in building['polygon_mesh']:
ring = []
for coord in surface:
key = (round(float(coord[0]), 6), round(float(coord[1]), 6), round(float(coord[2]), 6))
# Find the nearest stored key (handles float drift)
if key not in vertex_index:
key = min(
local_idx.keys(),
key=lambda k: (k[0]-key[0])**2 + (k[1]-key[1])**2 + (k[2]-key[2])**2
)
ring.append(vertex_index[key])
boundaries.append([ring])
city_objects[f"bag_{bid}"] = {
"type": "Building",
"geometry": [{
"type": "Solid",
"lod": "2",
"boundaries": [boundaries]
}],
"attributes": building.get("attributes", {})
}
cityjson = {
"type": "CityJSON",
"version": "1.1",
"metadata": {
"referenceSystem": f"https://www.opengis.net/def/crs/EPSG/0/{epsg_code}"
},
"CityObjects": city_objects,
"vertices": all_vertices,
"alignment_info": {
"mean_residual_m": round(self.mean_residual, 3),
"n_anchor_pairs": self.n_anchors,
"alpha": self.alpha
}
}
with open(out_path, 'w', encoding='utf-8') as f:
json.dump(cityjson, f, indent=2)
size_mb = os.path.getsize(out_path) / 1e6
self.logger.info(
f"[RigidAligner] Aligned CityJSON written: {out_path} "
f"({len(city_objects)} buildings, {size_mb:.1f} MB)"
)