py4dgeo 0.7.0__cp313-cp313-win_amd64.whl

py4dgeo/m3c2ep.py

@@ -0,0 +1,855 @@
+import numpy as np
+import py4dgeo
+import math
+import time
+import scipy.stats as sstats
+import multiprocessing as mp
+import laspy
+
+from py4dgeo.epoch import Epoch
+from py4dgeo.util import (
+    as_double_precision,
+    Py4DGeoError,
+)
+
+from py4dgeo import M3C2
+
+import warnings
+
+warnings.filterwarnings("ignore")
+
+try:
+    from tqdm import tqdm
+except ImportError:
+    tqdm = None
+
+default_tfM = np.array([[1, 0, 0, 0], [0, 1, 0, 0], [0, 0, 1, 0]])
+
+
+class M3C2EP(M3C2):
+    def __init__(
+        self,
+        tfM: np.ndarray = default_tfM,
+        Cxx: np.ndarray = np.zeros((12, 12)),
+        refPointMov: np.ndarray = np.array([0, 0, 0]),
+        perform_trans: bool = True,
+        **kwargs,
+    ):
+        """An M3C2-EP implementation
+        that push the limits of 3D topographic point cloud change detection by error propagation.
+        The algorithm needs an alignment covariance matrix of shape 12 x 12, an affine transformation matrix
+        of shape 3 x 4 and a reduction point  (x0,y0,z0) (rotation origin, 3 parameters) obtained from
+        aligning the two point clouds. The formula of the transformation see in user docs.
+        The transformation can be set by a boolean flag 'perform_trans' and is performed by default.
+        """
+        assert tfM.shape == (3, 4)
+        assert refPointMov.shape == (3,)
+        assert Cxx.shape == (12, 12)
+
+        self.tfM = tfM
+        self.Cxx = Cxx
+        self.refPointMov = refPointMov
+        self.perform_trans = perform_trans
+        super().__init__(**kwargs)
+
+    def calculate_distances(self, epoch1, epoch2):
+        print(self.name + " running")
+        """Calculate the distances between two epochs"""
+
+        if not isinstance(self.cyl_radius, float):
+            raise Py4DGeoError(
+                f"{self.name} requires exactly one cylinder radius to be given"
+            )
+
+        epoch1.build_kdtree()
+        epoch2.build_kdtree()
+
+        p1_coords = epoch1.cloud
+        p1_positions = epoch1.scanpos_id
+        p2_coords = epoch2.cloud
+        p2_positions = epoch2.scanpos_id
+
+        # set default M3C2Meta
+        M3C2Meta = {"searchrad": 0.5, "maxdist": 3, "minneigh": 5, "maxneigh": 100000}
+        M3C2Meta["searchrad"] = self.cyl_radius
+        M3C2Meta["maxdist"] = self.max_distance
+
+        M3C2Meta["spInfos"] = [epoch1.scanpos_info, epoch2.scanpos_info]
+        M3C2Meta["tfM"] = self.tfM
+        M3C2Meta["Cxx"] = self.Cxx
+        M3C2Meta["redPoint"] = self.refPointMov
+
+        refPointMov = self.refPointMov
+        tfM = self.tfM
+
+        # transform p2
+        if self.perform_trans:
+            p2_coords = p2_coords - refPointMov
+            p2_coords = np.dot(tfM[:3, :3], p2_coords.T).T + tfM[:, 3] + refPointMov
+
+        # load query points
+        query_coords = self.corepoints
+        query_norms = self.directions()
+
+        # Repeat normals to shape of corepoints when the user explicitly provided one direction
+        if query_norms.shape[0] == 1:
+            query_norms = query_norms.repeat(self.corepoints.shape[0], axis=0)
+
+        if query_norms is None:
+            raise Py4DGeoError("Core point point cloud needs normals set. Exiting.")
+            exit(-1)
+        subsample = False
+        if subsample:
+            sub_idx = np.random.choice(np.arange(0, query_coords.shape[0]), 2000)
+            query_coords = query_coords[sub_idx]
+            query_norms = query_norms[sub_idx]
+
+        NUM_THREADS = 4
+        NUM_BLOCKS = 16
+
+        query_coords_subs = np.array_split(query_coords, NUM_BLOCKS)
+        query_norms_subs = np.array_split(query_norms, NUM_BLOCKS)
+
+        # start mp
+        manager = mp.Manager()
+        return_dict = manager.dict()
+
+        # prepare shared memory
+        p1_coords_shm = mp.shared_memory.SharedMemory(
+            create=True, size=p1_coords.nbytes
+        )
+        p1_coords_sha = np.ndarray(
+            p1_coords.shape, dtype=p1_coords.dtype, buffer=p1_coords_shm.buf
+        )
+        p1_coords_sha[:] = p1_coords[:]
+        p2_coords_shm = mp.shared_memory.SharedMemory(
+            create=True, size=p2_coords.nbytes
+        )
+        p2_coords_sha = np.ndarray(
+            p2_coords.shape, dtype=p2_coords.dtype, buffer=p2_coords_shm.buf
+        )
+        p2_coords_sha[:] = p2_coords[:]
+
+        max_dist = M3C2Meta["maxdist"]
+        search_radius = M3C2Meta["searchrad"]
+        effective_search_radius = math.hypot(max_dist, search_radius)
+
+        # Querying neighbours
+        pbarQueue = mp.Queue()
+        pbarProc = mp.Process(
+            target=updatePbar, args=(query_coords.shape[0], pbarQueue, NUM_THREADS)
+        )
+        pbarProc.start()
+        procs = []
+
+        last_started_idx = -1
+        running_ps = []
+        while True:
+            if len(running_ps) < NUM_THREADS:
+                last_started_idx += 1
+                if last_started_idx < len(query_coords_subs):
+                    curr_subs = query_coords_subs[last_started_idx]
+                    p1_idx = radius_search(epoch1, curr_subs, effective_search_radius)
+                    p2_idx = radius_search(epoch2, curr_subs, effective_search_radius)
+
+                    p = mp.Process(
+                        target=process_corepoint_list,
+                        args=(
+                            curr_subs,
+                            query_norms_subs[last_started_idx],
+                            p1_idx,
+                            p1_coords_shm.name,
+                            p1_coords.shape,
+                            p1_positions,
+                            p2_idx,
+                            p2_coords_shm.name,
+                            p2_coords.shape,
+                            p2_positions,
+                            M3C2Meta,
+                            last_started_idx,
+                            return_dict,
+                            pbarQueue,
+                        ),
+                    )
+                    procs.append(p)
+
+                    procs[last_started_idx].start()
+                    running_ps.append(last_started_idx)
+                else:
+                    break
+            for running_p in running_ps:
+                if not procs[running_p].is_alive():
+                    running_ps.remove(running_p)
+            time.sleep(1)
+
+        for p in procs:
+            p.join()
+
+        pbarQueue.put((0, 0))
+        pbarProc.terminate()
+
+        p1_coords_shm.close()
+        p1_coords_shm.unlink()
+        p2_coords_shm.close()
+        p2_coords_shm.unlink()
+
+        out_attrs = {
+            key: (
+                np.empty((query_coords.shape[0], 3, 3), dtype=val.dtype)
+                if key == "m3c2_cov1" or key == "m3c2_cov2"
+                else np.empty(query_coords.shape[0], dtype=val.dtype)
+            )
+            for key, val in return_dict[0].items()
+        }
+        for key in out_attrs:
+            curr_start = 0
+            for i in range(NUM_BLOCKS):
+                curr_len = return_dict[i][key].shape[0]
+                out_attrs[key][curr_start : curr_start + curr_len] = return_dict[i][key]
+                curr_start += curr_len
+
+        distances = out_attrs["val"]
+        cov1 = out_attrs["m3c2_cov1"]
+        cov2 = out_attrs["m3c2_cov2"]
+        unc = {
+            "lodetection": out_attrs["lod_new"],
+            "spread1": out_attrs["m3c2_spread1"],
+            "num_samples1": out_attrs["m3c2_n1"],
+            "spread2": out_attrs["m3c2_spread2"],
+            "num_samples2": out_attrs["m3c2_n2"],
+        }
+
+        unc_list = []
+        cov_list = []
+        for i in range(unc["lodetection"].shape[0]):
+            unc_item = (
+                unc["lodetection"][i],
+                unc["spread1"][i],
+                unc["num_samples1"][i],
+                unc["spread2"][i],
+                unc["num_samples2"][i],
+            )
+            unc_list.append(unc_item)
+            cov_item = (cov1[i], cov2[i])
+            cov_list.append(cov_item)
+
+        uncertainties = np.array(
+            unc_list,
+            dtype=[
+                ("lodetection", "f8"),
+                ("spread1", "f8"),
+                ("num_samples1", "i8"),
+                ("spread2", "f8"),
+                ("num_samples2", "i8"),
+            ],
+        )
+        covariance = np.array(
+            cov_list, dtype=[("cov1", "f8", (3, 3)), ("cov2", "f8", (3, 3))]
+        )
+        print(self.name + " end")
+        return distances, uncertainties, covariance
+
+    @property
+    def name(self):
+        return "M3C2EP"
+
+
+def updatePbar(total, queue, maxProc):
+    desc = "Processing core points"
+    pCount = 0
+    if tqdm is None:
+        pbar = None
+    else:
+        pbar = tqdm(
+            total=total,
+            ncols=100,
+            desc=desc + " (%02d/%02d Process(es))" % (pCount, maxProc),
+        )
+
+    while True:
+        inc, process = queue.get()
+        if pbar is not None:
+            pbar.update(inc)
+            if process != 0:
+                pCount += process
+                pbar.set_description(
+                    desc + " (%02d/%02d Process(es))" % (pCount, maxProc)
+                )
+
+
+eijk = np.zeros((3, 3, 3))
+eijk[0, 1, 2] = eijk[1, 2, 0] = eijk[2, 0, 1] = 1
+eijk[0, 2, 1] = eijk[2, 1, 0] = eijk[1, 0, 2] = -1
+
+dij = np.zeros((3, 3))
+dij[0, 0] = dij[1, 1] = dij[2, 2] = 1
+
+n = np.zeros((3,))
+poa_pts = np.zeros((3, 100))
+path_opt = np.einsum_path(
+    "mi, ijk, j, kn -> mn", dij, eijk, n, poa_pts, optimize="optimal"
+)
+
+
+def getAlongAcrossSqBatch(pts, poa, n):
+    pts_poa = pts - poa[:, np.newaxis]
+    alongs = n.dot(pts_poa)
+    poa_pts = poa[:, np.newaxis] - pts
+    crosses = np.einsum(
+        "mi, ijk, j, kn -> mn", dij, eijk, n, poa_pts, optimize=path_opt[0]
+    )
+    across2 = np.einsum("ij, ij -> j", crosses, crosses)
+    return (alongs, across2)
+
+
+def get_local_mean_and_Cxx_nocorr(
+    Cxx, tfM, origins, redPoint, sigmas, curr_pts, curr_pos, epoch, tf=True
+):
+    nPts = curr_pts.shape[0]
+    A = np.tile(np.eye(3), (nPts, 1))
+    ATP = np.zeros((3, 3 * nPts))
+    tfM = tfM if tf else np.array([[1, 0, 0, 0], [0, 1, 0, 0], [0, 0, 1, 0]])
+    dx = np.zeros((nPts,), dtype=np.float64)
+    dy = np.zeros((nPts,), dtype=np.float64)
+    dz = np.zeros((nPts,), dtype=np.float64)
+    rrange = np.zeros((nPts,), dtype=np.float64)
+    sinscan = np.zeros((nPts,), dtype=np.float64)
+    cosscan = np.zeros((nPts,), dtype=np.float64)
+    cosyaw = np.zeros((nPts,), dtype=np.float64)
+    sinyaw = np.zeros((nPts,), dtype=np.float64)
+    sigmaRange = np.zeros((nPts,), dtype=np.float64)
+    sigmaYaw = np.zeros((nPts,), dtype=np.float64)
+    sigmaScan = np.zeros((nPts,), dtype=np.float64)
+
+    for scanPosId in np.unique(curr_pos):
+        scanPos = np.array(origins[scanPosId - 1, :])
+        scanPosPtsIdx = curr_pos == scanPosId
+
+        dd = curr_pts[scanPosPtsIdx, :] - scanPos[np.newaxis, :]
+        dlx, dly, dlz = dd[:, 0], dd[:, 1], dd[:, 2]
+        yaw = np.arctan2(dly, dlx)
+        planar_dist = np.hypot(dlx, dly)
+        scan = np.pi / 2 - np.arctan(dlz / planar_dist)
+        rrange[scanPosPtsIdx] = np.hypot(planar_dist, dlz)
+        sinscan[scanPosPtsIdx] = np.sin(scan)
+        cosscan[scanPosPtsIdx] = np.cos(scan)
+        sinyaw[scanPosPtsIdx] = np.sin(yaw)
+        cosyaw[scanPosPtsIdx] = np.cos(yaw)
+
+        dr = curr_pts[scanPosPtsIdx, :] - redPoint
+        dx[scanPosPtsIdx] = dr[:, 0]
+        dy[scanPosPtsIdx] = dr[:, 1]
+        dz[scanPosPtsIdx] = dr[:, 2]
+
+        sigmaRange[scanPosPtsIdx] = np.array(
+            np.sqrt(
+                sigmas[scanPosId - 1][0] ** 2
+                + sigmas[scanPosId - 1][1] * 1e-6 * rrange[scanPosPtsIdx] ** 2
+            )
+        )  # a + b*d
+        sigmaYaw[scanPosPtsIdx] = np.array(sigmas[scanPosId - 1][2])
+        sigmaScan[scanPosPtsIdx] = np.array(sigmas[scanPosId - 1][3])
+
+    if tf:
+        SigmaXiXj = (
+            dx**2 * Cxx[0, 0]
+            + 2 * dx * dy * Cxx[0, 1]  # a11a11
+            + dy**2 * Cxx[1, 1]  # a11a12
+            + 2 * dy * dz * Cxx[1, 2]  # a12a12
+            + dz**2 * Cxx[2, 2]  # a12a13
+            + 2 * dz * dx * Cxx[0, 2]  # a13a13
+            + 2  # a11a13
+            * (dx * Cxx[0, 9] + dy * Cxx[1, 9] + dz * Cxx[2, 9])  # a11tx  # a12tx
+            + Cxx[9, 9]  # a13tx
+        )  # txtx
+
+        SigmaYiYj = (
+            dx**2 * Cxx[3, 3]
+            + 2 * dx * dy * Cxx[3, 4]  # a21a21
+            + dy**2 * Cxx[4, 4]  # a21a22
+            + 2 * dy * dz * Cxx[4, 5]  # a22a22
+            + dz**2 * Cxx[5, 5]  # a22a23
+            + 2 * dz * dx * Cxx[3, 5]  # a23a23
+            + 2  # a21a23
+            * (dx * Cxx[3, 10] + dy * Cxx[4, 10] + dz * Cxx[5, 10])  # a21ty  # a22ty
+            + Cxx[10, 10]  # a23ty
+        )  # tyty
+
+        SigmaZiZj = (
+            dx**2 * Cxx[6, 6]
+            + 2 * dx * dy * Cxx[6, 7]  # a31a31
+            + dy**2 * Cxx[7, 7]  # a31a32
+            + 2 * dy * dz * Cxx[7, 8]  # a32a32
+            + dz**2 * Cxx[8, 8]  # a32a33
+            + 2 * dz * dx * Cxx[6, 8]  # a33a33
+            + 2  # a31a33
+            * (dx * Cxx[6, 11] + dy * Cxx[7, 11] + dz * Cxx[8, 11])  # a31tz  # a32tz
+            + Cxx[11, 11]  # a33tz
+        )  # tztz
+
+        SigmaXiYj = (
+            Cxx[9, 10]
+            + dx * Cxx[0, 10]  # txty
+            + dy * Cxx[1, 10]  # a11ty
+            + dz * Cxx[2, 10]  # a12ty
+            + dx  # a13ty
+            * (Cxx[3, 9] + Cxx[0, 3] * dx + Cxx[1, 3] * dy + Cxx[2, 3] * dz)
+            + dy * (Cxx[4, 9] + Cxx[0, 4] * dx + Cxx[1, 4] * dy + Cxx[2, 4] * dz)
+            + dz * (Cxx[5, 9] + Cxx[0, 5] * dx + Cxx[1, 5] * dy + Cxx[2, 5] * dz)
+        )
+
+        SigmaXiZj = (
+            Cxx[9, 11]
+            + dx * Cxx[0, 11]  # txtz
+            + dy * Cxx[1, 11]  # a11tz
+            + dz * Cxx[2, 11]  # a12tz
+            + dx  # a13tz
+            * (Cxx[6, 9] + Cxx[0, 6] * dx + Cxx[1, 6] * dy + Cxx[2, 6] * dz)
+            + dy * (Cxx[7, 9] + Cxx[0, 7] * dx + Cxx[1, 7] * dy + Cxx[2, 7] * dz)
+            + dz * (Cxx[8, 9] + Cxx[0, 8] * dx + Cxx[1, 8] * dy + Cxx[2, 8] * dz)
+        )
+
+        SigmaYiZj = (
+            Cxx[10, 11]
+            + dx * Cxx[6, 10]  # tytz
+            + dy * Cxx[7, 10]  # a21tx
+            + dz * Cxx[8, 10]  # a22tx
+            + dx  # a23tx
+            * (Cxx[3, 11] + Cxx[3, 6] * dx + Cxx[3, 7] * dy + Cxx[3, 8] * dz)
+            + dy * (Cxx[4, 11] + Cxx[4, 6] * dx + Cxx[4, 7] * dy + Cxx[4, 8] * dz)
+            + dz * (Cxx[5, 11] + Cxx[5, 6] * dx + Cxx[5, 7] * dy + Cxx[5, 8] * dz)
+        )
+        C11 = np.sum(SigmaXiXj)  # sum over all j
+        C12 = np.sum(SigmaXiYj)  # sum over all j
+        C13 = np.sum(SigmaXiZj)  # sum over all j
+        C22 = np.sum(SigmaYiYj)  # sum over all j
+        C23 = np.sum(SigmaYiZj)  # sum over all j
+        C33 = np.sum(SigmaZiZj)  # sum over all j
+        local_Cxx = np.array([[C11, C12, C13], [C12, C22, C23], [C13, C23, C33]])
+    else:
+        local_Cxx = np.zeros((3, 3))
+
+    C11p = (
+        (
+            tfM[0, 0] * cosyaw * sinscan
+            + tfM[0, 1] * sinyaw * sinscan  # dX/dRange - measurements
+            + tfM[0, 2] * cosscan
+        )
+        ** 2
+        * sigmaRange**2
+        + (
+            -1 * tfM[0, 0] * rrange * sinyaw * sinscan
+            + tfM[0, 1] * rrange * cosyaw * sinscan  # dX/dYaw
+        )
+        ** 2
+        * sigmaYaw**2
+        + (
+            tfM[0, 0] * rrange * cosyaw * cosscan
+            + tfM[0, 1] * rrange * sinyaw * cosscan  # dX/dScan
+            + -1 * tfM[0, 2] * rrange * sinscan
+        )
+        ** 2
+        * sigmaScan**2
+    )
+
+    C12p = (
+        (
+            tfM[1, 0] * cosyaw * sinscan
+            + tfM[1, 1] * sinyaw * sinscan  # dY/dRange - measurements
+            + tfM[1, 2] * cosscan
+        )
+        * (
+            tfM[0, 0] * cosyaw * sinscan
+            + tfM[0, 1] * sinyaw * sinscan  # dX/dRange - measurements
+            + tfM[0, 2] * cosscan
+        )
+        * sigmaRange**2
+        + (
+            -1 * tfM[1, 0] * rrange * sinyaw * sinscan
+            + tfM[1, 1] * rrange * cosyaw * sinscan  # dY/dYaw
+        )
+        * (
+            -1 * tfM[0, 0] * rrange * sinyaw * sinscan
+            + tfM[0, 1] * rrange * cosyaw * sinscan  # dX/dYaw
+        )
+        * sigmaYaw**2
+        + (
+            tfM[0, 0] * rrange * cosyaw * cosscan
+            + tfM[0, 1] * rrange * sinyaw * cosscan  # dX/dScan
+            + -1 * tfM[0, 2] * rrange * sinscan
+        )
+        * (
+            tfM[1, 0] * rrange * cosyaw * cosscan
+            + tfM[1, 1] * rrange * sinyaw * cosscan  # dY/dScan
+            + -1 * tfM[1, 2] * rrange * sinscan
+        )
+        * sigmaScan**2
+    )
+
+    C22p = (
+        (
+            tfM[1, 0] * cosyaw * sinscan
+            + tfM[1, 1] * sinyaw * sinscan  # dY/dRange - measurements
+            + tfM[1, 2] * cosscan
+        )
+        ** 2
+        * sigmaRange**2
+        + (
+            -1 * tfM[1, 0] * rrange * sinyaw * sinscan
+            + tfM[1, 1] * rrange * cosyaw * sinscan  # dY/dYaw
+        )
+        ** 2
+        * sigmaYaw**2
+        + (
+            tfM[1, 0] * rrange * cosyaw * cosscan
+            + tfM[1, 1] * rrange * sinyaw * cosscan  # dY/dScan
+            + -1 * tfM[1, 2] * rrange * sinscan
+        )
+        ** 2
+        * sigmaScan**2
+    )
+
+    C23p = (
+        (
+            tfM[1, 0] * cosyaw * sinscan
+            + tfM[1, 1] * sinyaw * sinscan  # dY/dRange - measurements
+            + tfM[1, 2] * cosscan
+        )
+        * (
+            tfM[2, 0] * cosyaw * sinscan
+            + tfM[2, 1] * sinyaw * sinscan  # dZ/dRange - measurements
+            + tfM[2, 2] * cosscan
+        )
+        * sigmaRange**2
+        + (
+            -1 * tfM[1, 0] * rrange * sinyaw * sinscan
+            + tfM[1, 1] * rrange * cosyaw * sinscan  # dY/dYaw
+        )
+        * (
+            -1 * tfM[2, 0] * rrange * sinyaw * sinscan
+            + tfM[2, 1] * rrange * cosyaw * sinscan  # dZ/dYaw
+        )
+        * sigmaYaw**2
+        + (
+            tfM[2, 0] * rrange * cosyaw * cosscan
+            + tfM[2, 1] * rrange * sinyaw * cosscan  # dZ/dScan
+            + -1 * tfM[2, 2] * rrange * sinscan
+        )
+        * (
+            tfM[1, 0] * rrange * cosyaw * cosscan
+            + tfM[1, 1] * rrange * sinyaw * cosscan  # dY/dScan
+            + -1 * tfM[1, 2] * rrange * sinscan
+        )
+        * sigmaScan**2
+    )
+
+    C33p = (
+        (
+            tfM[2, 0] * cosyaw * sinscan
+            + tfM[2, 1] * sinyaw * sinscan  # dZ/dRange - measurements
+            + tfM[2, 2] * cosscan
+        )
+        ** 2
+        * sigmaRange**2
+        + (
+            -1 * tfM[2, 0] * rrange * sinyaw * sinscan
+            + tfM[2, 1] * rrange * cosyaw * sinscan  # dZ/dYaw
+        )
+        ** 2
+        * sigmaYaw**2
+        + (
+            tfM[2, 0] * rrange * cosyaw * cosscan
+            + tfM[2, 1] * rrange * sinyaw * cosscan  # dZ/dScan
+            + -1 * tfM[2, 2] * rrange * sinscan
+        )
+        ** 2
+        * sigmaScan**2
+    )
+
+    C13p = (
+        (
+            tfM[2, 0] * cosyaw * sinscan
+            + tfM[2, 1] * sinyaw * sinscan  # dZ/dRange - measurements
+            + tfM[2, 2] * cosscan
+        )
+        * (
+            tfM[0, 0] * cosyaw * sinscan
+            + tfM[0, 1] * sinyaw * sinscan  # dX/dRange - measurements
+            + tfM[0, 2] * cosscan
+        )
+        * sigmaRange**2
+        + (
+            -1 * tfM[2, 0] * rrange * sinyaw * sinscan
+            + tfM[2, 1] * rrange * cosyaw * sinscan  # dZ/dYaw
+        )
+        * (
+            -1 * tfM[0, 0] * rrange * sinyaw * sinscan
+            + tfM[0, 1] * rrange * cosyaw * sinscan  # dX/dYaw
+        )
+        * sigmaYaw**2
+        + (
+            tfM[2, 0] * rrange * cosyaw * cosscan
+            + tfM[2, 1] * rrange * sinyaw * cosscan  # dZ/dScan
+            + -1 * tfM[2, 2] * rrange * sinscan
+        )
+        * (
+            tfM[0, 0] * rrange * cosyaw * cosscan
+            + tfM[0, 1] * rrange * sinyaw * cosscan  # dX/dScan
+            + -1 * tfM[0, 2] * rrange * sinscan
+        )
+        * sigmaScan**2
+    )
+    local_Cxx[0, 0] += np.sum(C11p)
+    local_Cxx[0, 1] += np.sum(C12p)
+    local_Cxx[0, 2] += np.sum(C13p)
+    local_Cxx[1, 0] += np.sum(C12p)
+    local_Cxx[1, 1] += np.sum(C22p)
+    local_Cxx[1, 2] += np.sum(C23p)
+    local_Cxx[2, 1] += np.sum(C23p)
+    local_Cxx[2, 0] += np.sum(C13p)
+    local_Cxx[2, 2] += np.sum(C33p)
+
+    # Get mean without correlation (averages out anyway, or something...)
+    for pii in range(nPts):
+        Cxx = np.array(
+            [
+                [C11p[pii], C12p[pii], C13p[pii]],
+                [C12p[pii], C22p[pii], C23p[pii]],
+                [C13p[pii], C23p[pii], C33p[pii]],
+            ]
+        )
+        if np.linalg.det(Cxx) == 0:
+            Cxx = np.eye(3)
+        Cix = np.linalg.inv(Cxx)
+        ATP[:, pii * 3 : (pii + 1) * 3] = Cix
+    N = np.dot(ATP, A)
+    Qxx = np.linalg.inv(N)  # can only have > 0 in main diagonal!
+    pts_m = curr_pts.mean(axis=0)
+    l = (curr_pts - pts_m).flatten()
+    mean = np.dot(Qxx, np.dot(ATP, l)) + pts_m
+
+    return mean, local_Cxx / nPts
+
+
+def process_corepoint_list(
+    corepoints,
+    corepoint_normals,
+    p1_idx,
+    p1_shm_name,
+    p1_size,
+    p1_positions,
+    p2_idx,
+    p2_shm_name,
+    p2_size,
+    p2_positions,
+    M3C2Meta,
+    idx,
+    return_dict,
+    pbarQueue,
+):
+    pbarQueue.put((0, 1))
+    p1_shm = mp.shared_memory.SharedMemory(name=p1_shm_name)
+    p2_shm = mp.shared_memory.SharedMemory(name=p2_shm_name)
+    p1_coords = np.ndarray(p1_size, dtype=np.float64, buffer=p1_shm.buf)
+    p2_coords = np.ndarray(p2_size, dtype=np.float64, buffer=p2_shm.buf)
+
+    max_dist = M3C2Meta["maxdist"]
+    search_radius = M3C2Meta["searchrad"]
+
+    M3C2_vals = np.full((corepoints.shape[0]), np.nan, dtype=np.float64)
+    M3C2_LoD = np.full((corepoints.shape[0]), np.nan, dtype=np.float64)
+    M3C2_N1 = np.full((corepoints.shape[0]), np.nan, dtype=np.int32)
+    M3C2_N2 = np.full((corepoints.shape[0]), np.nan, dtype=np.int32)
+
+    M3C2_spread1 = np.full((corepoints.shape[0]), np.nan, dtype=np.float64)
+    M3C2_spread2 = np.full((corepoints.shape[0]), np.nan, dtype=np.float64)
+    M3C2_cov1 = np.full((corepoints.shape[0], 3, 3), np.nan, dtype=np.float64)
+    M3C2_cov2 = np.full((corepoints.shape[0], 3, 3), np.nan, dtype=np.float64)
+
+    for cp_idx, p1_neighbours in enumerate(p1_idx):
+        n = corepoint_normals[cp_idx]
+        p1_curr_pts = p1_coords[p1_neighbours, :]
+        along1, acrossSq1 = getAlongAcrossSqBatch(p1_curr_pts.T, corepoints[cp_idx], n)
+        p1_curr_pts = p1_curr_pts[
+            np.logical_and(np.abs(along1) <= max_dist, acrossSq1 <= search_radius**2),
+            :,
+        ]
+        p1_scanPos = p1_positions[p1_neighbours]
+        p1_scanPos = p1_scanPos[
+            np.logical_and(np.abs(along1) <= max_dist, acrossSq1 <= search_radius**2)
+        ]
+        if p1_curr_pts.shape[0] < M3C2Meta["minneigh"]:
+            pbarQueue.put((1, 0))  # point processed
+            M3C2_N1[cp_idx] = p1_curr_pts.shape[0]
+            continue
+        elif p1_curr_pts.shape[0] > M3C2Meta["maxneigh"]:
+            p1_curr_pts = p1_curr_pts[np.argsort(acrossSq1[: M3C2Meta["maxneigh"]])]
+            p1_scanPos = p1_scanPos[np.argsort(acrossSq1[: M3C2Meta["maxneigh"]])]
+
+        Cxx = M3C2Meta["Cxx"]
+        tfM = M3C2Meta["tfM"]
+        origins = np.array([SP["origin"] for SP in M3C2Meta["spInfos"][0]])
+        redPoint = M3C2Meta["redPoint"]
+        sigmas = np.array(
+            [
+                [
+                    SP["sigma_range"],
+                    SP["sigma_range"],
+                    SP["sigma_scan"],
+                    SP["sigma_yaw"],
+                ]
+                for SP in M3C2Meta["spInfos"][0]
+            ]
+        )
+
+        p1_weighted_CoG, p1_local_Cxx = get_local_mean_and_Cxx_nocorr(
+            Cxx,
+            tfM,
+            origins,
+            redPoint,
+            sigmas,
+            p1_curr_pts,
+            p1_scanPos,
+            epoch=0,
+            tf=False,
+        )  # only one dataset has been transformed
+        along1_var = np.var(
+            along1[
+                np.logical_and(
+                    np.abs(along1) <= max_dist, acrossSq1 <= search_radius**2
+                )
+            ]
+        )
+
+        p2_neighbours = p2_idx[cp_idx]
+        p2_curr_pts = p2_coords[p2_neighbours, :]
+        along2, acrossSq2 = getAlongAcrossSqBatch(p2_curr_pts.T, corepoints[cp_idx], n)
+        p2_curr_pts = p2_curr_pts[
+            np.logical_and(np.abs(along2) <= max_dist, acrossSq2 <= search_radius**2),
+            :,
+        ]
+        p2_scanPos = p2_positions[p2_neighbours]
+        p2_scanPos = p2_scanPos[
+            np.logical_and(np.abs(along2) <= max_dist, acrossSq2 <= search_radius**2)
+        ]
+        if p2_curr_pts.shape[0] < M3C2Meta["minneigh"]:
+            pbarQueue.put((1, 0))  # point processed
+            M3C2_N2[cp_idx] = p2_curr_pts.shape[0]
+            continue
+        elif p2_curr_pts.shape[0] > M3C2Meta["maxneigh"]:
+            p2_curr_pts = p2_curr_pts[np.argsort(acrossSq2[: M3C2Meta["maxneigh"]])]
+            p2_scanPos = p2_scanPos[np.argsort(acrossSq2[: M3C2Meta["maxneigh"]])]
+
+        origins = np.array([SP["origin"] for SP in M3C2Meta["spInfos"][1]])
+        sigmas = np.array(
+            [
+                [
+                    SP["sigma_range"],
+                    SP["sigma_range"],
+                    SP["sigma_scan"],
+                    SP["sigma_yaw"],
+                ]
+                for SP in M3C2Meta["spInfos"][1]
+            ]
+        )
+        p2_weighted_CoG, p2_local_Cxx = get_local_mean_and_Cxx_nocorr(
+            Cxx,
+            tfM,
+            origins,
+            redPoint,
+            sigmas,
+            p2_curr_pts,
+            p2_scanPos,
+            epoch=1,
+            tf=True,
+        )
+        along2_var = np.var(
+            along2[
+                np.logical_and(
+                    np.abs(along2) <= max_dist, acrossSq2 <= search_radius**2
+                )
+            ]
+        )
+
+        p1_CoG = p1_weighted_CoG
+        p2_CoG = p2_weighted_CoG
+
+        p1_CoG_Cxx = p1_local_Cxx
+        p2_CoG_Cxx = p2_local_Cxx
+
+        p1_p2_CoG_Cxx = np.zeros((6, 6))
+        p1_p2_CoG_Cxx[0:3, 0:3] = p1_CoG_Cxx
+        p1_p2_CoG_Cxx[3:6, 3:6] = p2_CoG_Cxx
+
+        M3C2_dist = n.dot(p1_CoG - p2_CoG)
+        F = np.hstack([-n, n])
+
+        M3C2_vals[cp_idx] = M3C2_dist
+
+        N1 = p1_curr_pts.shape[0]
+        N2 = p2_curr_pts.shape[0]
+
+        sigmaD = p1_CoG_Cxx + p2_CoG_Cxx
+
+        p = 3  # three dimensional
+        Tsqalt = n.T.dot(np.linalg.inv(sigmaD)).dot(n)
+
+        M3C2_LoD[cp_idx] = np.sqrt(sstats.chi2.ppf(0.95, p) / Tsqalt)
+        M3C2_N1[cp_idx] = N1
+        M3C2_N2[cp_idx] = N2
+
+        # add M3C2 spreads
+        normal = n[np.newaxis, :]
+        M3C2_spread1[cp_idx] = np.sqrt(
+            np.matmul(np.matmul(normal, p1_CoG_Cxx), normal.T)
+        )
+        M3C2_spread2[cp_idx] = np.sqrt(
+            np.matmul(np.matmul(normal, p2_CoG_Cxx), normal.T)
+        )
+        M3C2_cov1[cp_idx] = p1_CoG_Cxx
+        M3C2_cov2[cp_idx] = p2_CoG_Cxx
+
+        pbarQueue.put((1, 0))  # point processed
+
+    return_dict[idx] = {
+        "lod_new": M3C2_LoD,
+        "val": M3C2_vals,
+        "m3c2_n1": M3C2_N1,
+        "m3c2_n2": M3C2_N2,
+        "m3c2_spread1": M3C2_spread1,
+        "m3c2_spread2": M3C2_spread2,
+        "m3c2_cov1": M3C2_cov1,
+        "m3c2_cov2": M3C2_cov2,
+    }
+    pbarQueue.put((0, -1))
+    p1_shm.close()
+    p2_shm.close()
+
+
+def radius_search(epoch: Epoch, query: np.ndarray, radius: float):
+    """Query the tree for neighbors within a radius r
+    :param query:
+        An array of points to query.
+        Array-like of shape (n_samples, 3) or query 1 sample point of shape (3,)
+    :type query: array
+    :param radius:
+        Rebuild the search tree even if it was already built before.
+    :type radius: float
+    """
+    if len(query.shape) == 1 and query.shape[0] == 3:
+        return [epoch.kdtree.radius_search(query, radius)]
+
+    if len(query.shape) == 2 and query.shape[1] == 3:
+        neighbors = []
+        for i in range(query.shape[0]):
+            q = query[i]
+            result = epoch.kdtree.radius_search(q, radius)
+            neighbors.append(result)
+        return neighbors
+
+    raise Py4DGeoError(
+        "Please ensure queries are array-like of shape (n_samples, 3)"
+        " or of shape (3,) to query 1 sample point!"
+    )
+
+    return None