reach_for_the_arcs

reach_for_the_arcs(U, S, rng_seed=3452, return_point_cloud=False, fine_tune_iters=3, batch_size=10000, num_rasterization_spheres=0, screening_weight=10.0, rasterization_resolution=None, max_points_per_sphere=3, n_local_searches=None, local_search_iters=20, local_search_t=0.01, tol=0.0001, clamp_value=np.inf, force_cpu=False, parallel=False, verbose=False)

Creates a mesh from a signed distance function (SDF) using the "Reach for the Arcs" method of S. Sellán, Y. Ren, C. Batty, and O. Stein [2024]. This works for polylines in 2D or triangle meshes in 3D.

Parameters:

Name	Type	Description	Default
U	(n_sdf,d) numpy array	points where the SDF is evaluated	required
S	(n_sdf,) numpy array	sdf(U)	required
rng_seed	int, optional (default 3452)	rng seed where random data is needed	3452
return_point_cloud	bool, optional (default False)	return the reconstructed point cloud normals or not	False
fine_tune_iters	int, optional (default 5)	how may iterations to do in the fine tuning step	3
batch_size	int, optional (default 1000)	how many points in one batch. Set to 0 to disable batching.	10000
num_rasterization_spheres	int, optional (default 0)	how many spheres to use at most in the rasterization step. Set to zero to use all spheres.	0
screening_weight	float, optional (default 0.1)	PSR screening weight	10.0
rasterization_resolution	int, optional (default 8*n_sdf**(1/d))	the resolution of the rasterization grid	None
max_points_per_sphere	int, optional (default 3)	How many points should there be at most per sphere. Set it to 1 to not add any additional points to any sphere when fine-tuning. Has to be at least 1 (this is a theoretical minimum). If set to larger than 1, spheres that the reconstructed surface intersects will have at most max_points_per_sphere added.	3
n_local_searches	int, optional (default 2*n_sdf**(1/d))	how many local searches to perform for each sphere in the locally make feasible step	None
local_search_iters	int, optional (default 20)	how many iterations to try for in the local search for each sphere in the locally make feasible step	20
local_search_t	double, optional (default 5e-3)	how far to move each point in the local search for each point in the locally make feasible step	0.01
tol	float, optional (default 1e-4)	tolerance for determining whether a point is inside a sphere	0.0001
clamp_value	float, optional (default np.inf)	value to which the SDF is clamped for clamped SDF reconstruction	inf
force_cpu	bool, optional (default False)	whether to force rasterization onto the CPU	False
parallel	bool, optional (default False)	whether to parallelize the algorithm or not	False
verbose	bool, optional (default False)	whether to output details on the algorithm when it's running	False

Returns:

Name	Type	Description
V	(n,d) numpy array	reconstructed vertices
F	(m,d) numpy array	reconstructed polyline / triangle mesh indices
P	(n_p,d) numpy array, if requested	reconstructed point cloud points
N	(n_p,d) numpy array, if requested	reconstructed point cloud normals

Source code in src/gpytoolbox/reach_for_the_arcs.py

def reach_for_the_arcs(U, S,
    rng_seed=3452,
    return_point_cloud=False,
    fine_tune_iters=3,
    batch_size=10000,
    num_rasterization_spheres=0,
    screening_weight=10.,
    rasterization_resolution=None,
    max_points_per_sphere=3,
    n_local_searches=None,
    local_search_iters=20,
    local_search_t=0.01,
    tol=1e-4,
    clamp_value=np.inf,
    force_cpu=False,
    parallel=False,
    verbose=False):
    """Creates a mesh from a signed distance function (SDF) using the
    "Reach for the Arcs" method of S. Sellán,  Y. Ren, C. Batty, and O. Stein [2024].
    This works for polylines in 2D or triangle meshes in 3D.

    Parameters
    ----------
    U : (n_sdf,d) numpy array
        points where the SDF is evaluated
    S : (n_sdf,) numpy array
        sdf(U)
    rng_seed : int, optional (default 3452)
        rng seed where random data is needed
    return_point_cloud : bool, optional (default False)
        return the reconstructed point cloud normals or not
    fine_tune_iters : int, optional (default 5)
        how may iterations to do in the fine tuning step
    batch_size : int, optional (default 1000)
        how many points in one batch. Set to 0 to disable batching.
    num_rasterization_spheres : int, optional (default 0)
        how many spheres to use at most in the rasterization step.
        Set to zero to use all spheres.
    screening_weight : float, optional (default 0.1)
        PSR screening weight
    rasterization_resolution : int, optional (default 8*n_sdf**(1/d))
        the resolution of the rasterization grid
    max_points_per_sphere : int, optional (default 3)
        How many points should there be at most per sphere.
        Set it to 1 to not add any additional points to any sphere when
        fine-tuning.
        Has to be at least 1 (this is a theoretical minimum).
        If set to larger than 1, spheres that the reconstructed surface
        intersects will have at most max_points_per_sphere added.
    n_local_searches : int, optional (default 2*n_sdf**(1/d))
        how many local searches to perform for each sphere in the locally make
        feasible step
    local_search_iters : int, optional (default 20)
        how many iterations to try for in the local search for each sphere in
        the locally make feasible step
    local_search_t : double, optional (default 5e-3)
        how far to move each point in the local search for each point in the
        locally make feasible step
    tol : float, optional (default 1e-4)
        tolerance for determining whether a point is inside a sphere
    clamp_value : float, optional (default np.inf)
        value to which the SDF is clamped for clamped SDF reconstruction
    force_cpu : bool, optional (default False)
        whether to force rasterization onto the CPU
    parallel : bool, optional (default False)
        whether to parallelize the algorithm or not
    verbose : bool, optional (default False)
        whether to output details on the algorithm when it's running


    Returns
    -------
    V : (n,d) numpy array
        reconstructed vertices
    F : (m,d) numpy array
        reconstructed polyline / triangle mesh indices
    P : (n_p,d) numpy array, if requested
        reconstructed point cloud points
    N : (n_p,d) numpy array, if requested
        reconstructed point cloud normals
    """

    d = U.shape[1]
    assert d==2 or d==3, "Only dimensions 2 and 3 supported."

    n_sdf = U.shape[0]

    # Pick default values if not supplied.
    if rasterization_resolution is None:
        rasterization_resolution = 64 * math.ceil(n_sdf**(1./d)/16.)
    if n_local_searches is None:
        n_local_searches = math.ceil(2. * n_sdf**(1./d))

    # RNG used to compute random numbers and new seeds during the method.
    rng = np.random.default_rng(seed=rng_seed)
    seed = lambda : rng.integers(0,np.iinfo(np.int32).max)

    # Resize the SDF points in U and the SDF samples in S so it's in [0,1]^d
    trans = np.min(U, axis=0)
    U = U - trans[None,:]
    scale = np.max(U)
    U /= scale
    S = S/scale
    clamp_value = clamp_value/scale


    if verbose:
        print(f" --- Starting Reach for the Arcs --- ")
        t0_total = time.time()

    if verbose:
        print(f"SDF to point cloud...")
        t0_sdf_to_point_cloud = time.time()

    # Split the SDF into positive and negative spheres?
    separate_inside_outside = True
    if separate_inside_outside:
        neg = S<0
        pos = np.logical_not(neg)
        pos,neg = np.nonzero(pos)[0], np.nonzero(neg)[0]
        U_pos, U_neg = U[pos,:], U[neg,:]
        S_pos, S_neg = S[pos], S[neg]
        if pos.size > 0:
            if verbose:
                print(f"  positive spheres")
            P_pos,N_pos,f_pos = _sdf_to_point_cloud(U_pos, S_pos,
                rng_seed=seed(),
                rasterization_resolution=rasterization_resolution,
                n_local_searches=n_local_searches,
                local_search_iters=local_search_iters,
                batch_size=batch_size,
                num_rasterization_spheres=num_rasterization_spheres,
                force_cpu=force_cpu,
                tol=tol, clamp_value=clamp_value,
                parallel=parallel, verbose=verbose)
        else:
            P_pos,N_pos,f_pos = None,None,None
        if neg.size>0:
            if verbose:
                print(f"  negative spheres")
            P_neg,N_neg,f_neg = _sdf_to_point_cloud(U_neg, S_neg,
                rng_seed=seed(),
                rasterization_resolution=rasterization_resolution,
                n_local_searches=n_local_searches,
                local_search_iters=local_search_iters,
                batch_size=batch_size,
                num_rasterization_spheres=num_rasterization_spheres,
                force_cpu=force_cpu,
                tol=tol, clamp_value=clamp_value,
                parallel=parallel, verbose=verbose)
        else:
            P_neg,N_neg,f_neg = None,None,None

        if P_pos is None or P_pos.size==0:
            P,N,f = P_neg,N_neg,neg[f_neg]
        elif P_neg is None or P_neg.size==0:
            P,N,f = P_pos,N_pos,pos[f_pos]
        else:
            P = np.concatenate((P_pos, P_neg), axis=0)
            N = np.concatenate((N_pos, N_neg), axis=0)
            f = np.concatenate((pos[f_pos], neg[f_neg]), axis=0)
    else:
        P,N,f = _sdf_to_point_cloud(U, S,
            rng_seed=seed(),
            rasterization_resolution=rasterization_resolution,
            n_local_searches=n_local_searches,
            local_search_iters=local_search_iters,
            batch_size=batch_size,
            force_cpu=force_cpu,
            tol=tol, clamp_value=clamp_value,
            parallel=parallel, verbose=verbose)

    if P is None or P.size==0:
        if verbose:
            print(f"Unable to find any point cloud point.")
        if return_point_cloud:
            return V, F, P, N
        else:
            return V, F

    if verbose:
        print(f"SDF to point cloud took {time.time()-t0_sdf_to_point_cloud}s")
        print("")

    if verbose:
        print(f"Fine tuning point cloud...")
        t0_fine_tune = time.time()

    P,N,f = _fine_tune_point_cloud(U, S, P, N, f,
        rng_seed=seed(),
        fine_tune_iters=fine_tune_iters,
        batch_size=batch_size,
        screening_weight=screening_weight,
        max_points_per_sphere=max_points_per_sphere,
        n_local_searches=n_local_searches,
        local_search_iters=local_search_iters,
        local_search_t=local_search_t,
        tol=tol,clamp_value=clamp_value,
        parallel=parallel, verbose=verbose)

    if verbose:
        print(f"Fine tuning point cloud took {time.time()-t0_fine_tune}s")
        print("")

    if verbose:
        print(f"Converting point cloud to mesh...")
        t0_point_cloud_to_mesh = time.time()

    V,F = point_cloud_to_mesh(P, N,
        method='PSR',
        psr_screening_weight=screening_weight,
        psr_outer_boundary_type="Neumann",
        verbose=False)

    if V is None or V.size==0:
        if verbose:
            print(f"Reconstructing point cloud failed.")
        if return_point_cloud: 
            if P is None or P.size==0:
                return V, F, P, N
            else:
                return V, F, scale*P+trans[None,:], N
        else:
            return V, F

    if verbose:
        print(f"Converting point cloud to mesh took {time.time()-t0_point_cloud_to_mesh}s")
        print("")

    if verbose:
        print(f" --- Finished Reach for the Arcs --- ")
        print(f"Total elapsed time: {time.time()-t0_total}s")

    # Undo recentering
    V = scale*V + trans[None,:]
    P = scale*P + trans[None,:]

    if return_point_cloud:
        return V, F, P, N
    else:
        return V, F