voxcity 0.6.2__py3-none-any.whl → 0.6.4__py3-none-any.whl

voxcity/geoprocessor/mesh.py

@@ -1,791 +1,791 @@
-import numpy as np
-import os
-import trimesh
-import matplotlib.colors as mcolors
-import matplotlib.cm as cm
-import matplotlib.pyplot as plt
-
-def create_voxel_mesh(voxel_array, class_id, meshsize=1.0, building_id_grid=None, mesh_type=None):
-    """
-    Create a 3D mesh from voxels preserving sharp edges, scaled by meshsize.
-    
-    This function converts a 3D voxel array into a triangulated mesh, where each voxel
-    face is converted into two triangles. The function preserves sharp edges between
-    different classes and handles special cases for buildings.
-    
-    Parameters
-    ----------
-    voxel_array : np.ndarray (3D)
-        The voxel array of shape (X, Y, Z) where each cell contains a class ID.
-        - 0: typically represents void/air
-        - -2: typically represents trees
-        - -3: typically represents buildings
-        Other values can represent different classes as defined by the application.
-    
-    class_id : int
-        The ID of the class to extract. Only voxels with this ID will be included
-        in the output mesh.
-    
-    meshsize : float, default=1.0
-        The real-world size of each voxel in meters, applied uniformly to x, y, and z
-        dimensions. Used to scale the output mesh to real-world coordinates.
-    
-    building_id_grid : np.ndarray (2D), optional
-        2D grid of building IDs with shape (X, Y). Only used when class_id=-3 (buildings).
-        Each cell contains a unique identifier for the building at that location.
-        This allows tracking which faces belong to which building.
-    
-    mesh_type : str, optional
-        Type of mesh to create, controlling which faces are included:
-        - None (default): create faces at boundaries between different classes
-        - 'building_solar': only create faces at boundaries between buildings (-3)
-                          and either void (0) or trees (-2). Useful for solar analysis
-                          where only exposed surfaces matter.
-
-    Returns
-    -------
-    mesh : trimesh.Trimesh or None
-        The resulting triangulated mesh for the given class_id. Returns None if no
-        voxels of the specified class are found.
-        
-        The mesh includes:
-        - vertices: 3D coordinates of each vertex
-        - faces: triangles defined by vertex indices
-        - face_normals: normal vectors for each face
-        - metadata: If class_id=-3, includes 'building_id' mapping faces to buildings
-    
-    Examples
-    --------
-    Basic usage for a simple voxel array:
-    >>> voxels = np.zeros((10, 10, 10))
-    >>> voxels[4:7, 4:7, 0:5] = 1  # Create a simple column
-    >>> mesh = create_voxel_mesh(voxels, class_id=1, meshsize=0.5)
-    
-    Creating a building mesh with IDs:
-    >>> building_ids = np.zeros((10, 10))
-    >>> building_ids[4:7, 4:7] = 1  # Mark building #1
-    >>> mesh = create_voxel_mesh(voxels, class_id=-3, 
-    ...                         building_id_grid=building_ids,
-    ...                         meshsize=1.0)
-    
-    Notes
-    -----
-    - The function creates faces only at boundaries between different classes or at
-      the edges of the voxel array.
-    - Each face is split into two triangles for compatibility with graphics engines.
-    - Face normals are computed to ensure correct lighting and rendering.
-    - For buildings (class_id=-3), building IDs are tracked to maintain building identity.
-    - The mesh preserves sharp edges, which is important for architectural visualization.
-    """
-    # Find voxels of the current class
-    voxel_coords = np.argwhere(voxel_array == class_id)
-
-    if building_id_grid is not None:
-        building_id_grid_flipud = np.flipud(building_id_grid)
-
-    if len(voxel_coords) == 0:
-        return None
-
-    # Define the 6 faces of a unit cube (local coordinates 0..1)
-    unit_faces = np.array([
-        # Front
-        [[0, 0, 1], [1, 0, 1], [1, 1, 1], [0, 1, 1]],
-        # Back
-        [[0, 0, 0], [0, 1, 0], [1, 1, 0], [1, 0, 0]],
-        # Right
-        [[1, 0, 0], [1, 1, 0], [1, 1, 1], [1, 0, 1]],
-        # Left
-        [[0, 0, 0], [0, 0, 1], [0, 1, 1], [0, 1, 0]],
-        # Top
-        [[0, 1, 0], [0, 1, 1], [1, 1, 1], [1, 1, 0]],
-        # Bottom
-        [[0, 0, 0], [1, 0, 0], [1, 0, 1], [0, 0, 1]]
-    ])
-
-    # Define face normals
-    face_normals = np.array([
-        [0, 0, 1],   # Front
-        [0, 0, -1],  # Back
-        [1, 0, 0],   # Right
-        [-1, 0, 0],  # Left
-        [0, 1, 0],   # Top
-        [0, -1, 0]   # Bottom
-    ])
-
-    vertices = []
-    faces = []
-    face_normals_list = []
-    building_ids = []  # List to store building IDs for each face
-
-    for x, y, z in voxel_coords:
-        # For buildings, get the building ID from the grid
-        building_id = None
-        if class_id == -3 and building_id_grid is not None:
-            building_id = building_id_grid_flipud[x, y]
-            
-        # Check each face of the current voxel
-        adjacent_coords = [
-            (x,   y,   z+1),  # Front
-            (x,   y,   z-1),  # Back
-            (x+1, y,   z),    # Right
-            (x-1, y,   z),    # Left
-            (x,   y+1, z),    # Top
-            (x,   y-1, z)     # Bottom
-        ]
-
-        # Only create faces where there's a transition based on mesh_type
-        for face_idx, adj_coord in enumerate(adjacent_coords):
-            try:
-                # If adj_coord is outside array bounds, it's a boundary => face is visible
-                if adj_coord[0] < 0 or adj_coord[1] < 0 or adj_coord[2] < 0:
-                    is_boundary = True
-                else:
-                    adj_value = voxel_array[adj_coord]
-                    
-                    if mesh_type == 'open_air' and class_id == -3:
-                        # For building_solar, only create faces at boundaries with void (0) or trees (-2)
-                        is_boundary = (adj_value == 0 or adj_value == -2)
-                    else:
-                        # Default behavior - create faces at any class change
-                        is_boundary = (adj_value == 0 or adj_value != class_id)
-            except IndexError:
-                # Out of range => boundary
-                is_boundary = True
-
-            if is_boundary:
-                # Local face in (0..1) for x,y,z, then shift by voxel coords
-                face_verts = (unit_faces[face_idx] + np.array([x, y, z])) * meshsize
-                current_vert_count = len(vertices)
-
-                vertices.extend(face_verts)
-                # Convert quad to two triangles
-                faces.extend([
-                    [current_vert_count, current_vert_count + 1, current_vert_count + 2],
-                    [current_vert_count, current_vert_count + 2, current_vert_count + 3]
-                ])
-                # Add face normals for both triangles
-                face_normals_list.extend([face_normals[face_idx], face_normals[face_idx]])
-                
-                # Store building ID for both triangles if this is a building
-                if class_id == -3 and building_id_grid is not None:
-                    building_ids.extend([building_id, building_id])
-
-    if not vertices:
-        return None
-
-    vertices = np.array(vertices)
-    faces = np.array(faces)
-    face_normals_list = np.array(face_normals_list)
-
-    # Create mesh
-    mesh = trimesh.Trimesh(
-        vertices=vertices,
-        faces=faces,
-        face_normals=face_normals_list
-    )
-
-    # Merge vertices that are at the same position
-    mesh.merge_vertices()
-    
-    # Ensure metadata dict exists
-    if not hasattr(mesh, 'metadata') or mesh.metadata is None:
-        mesh.metadata = {}
-
-    # Store intended per-triangle normals to avoid reliance on auto-computed normals
-    mesh.metadata['provided_face_normals'] = face_normals_list
-
-    # Add building IDs as metadata for buildings
-    if class_id == -3 and building_id_grid is not None and building_ids:
-        mesh.metadata['building_id'] = np.array(building_ids)
-
-    return mesh
-
-def create_sim_surface_mesh(sim_grid, dem_grid,
-                            meshsize=1.0, z_offset=1.5,
-                            cmap_name='viridis',
-                            vmin=None, vmax=None):
-    """
-    Create a colored planar surface mesh from simulation data, positioned above a DEM.
-    
-    This function generates a 3D visualization mesh for 2D simulation results (like
-    Green View Index, solar radiation, etc.). The mesh is positioned above the Digital
-    Elevation Model (DEM) by a specified offset, and colored according to the simulation
-    values using a matplotlib colormap.
-    
-    Parameters
-    ----------
-    sim_grid : 2D np.ndarray
-        2D array of simulation values (e.g., Green View Index, solar radiation).
-        NaN values in this grid will be skipped in the output mesh.
-        The grid should be oriented with north at the top.
-    
-    dem_grid : 2D np.ndarray
-        2D array of ground elevations in meters. Must have the same shape as sim_grid.
-        Used to position the visualization mesh at the correct height above terrain.
-    
-    meshsize : float, default=1.0
-        Size of each cell in meters. Applied uniformly to x and y dimensions.
-        Determines the resolution of the output mesh.
-    
-    z_offset : float, default=1.5
-        Additional height offset in meters added to dem_grid elevations.
-        Used to position the visualization above ground level for better visibility.
-    
-    cmap_name : str, default='viridis'
-        Matplotlib colormap name used for coloring the mesh based on sim_grid values.
-        Common options:
-        - 'viridis': Default, perceptually uniform, colorblind-friendly
-        - 'RdYlBu': Red-Yellow-Blue, good for diverging data
-        - 'jet': Rainbow colormap (not recommended for scientific visualization)
-    
-    vmin : float, optional
-        Minimum value for color mapping. If None, uses min of sim_grid (excluding NaN).
-        Used to control the range of the colormap.
-    
-    vmax : float, optional
-        Maximum value for color mapping. If None, uses max of sim_grid (excluding NaN).
-        Used to control the range of the colormap.
-    
-    Returns
-    -------
-    mesh : trimesh.Trimesh or None
-        A single mesh containing one colored square face (two triangles) per non-NaN cell.
-        Returns None if there are no valid (non-NaN) cells in sim_grid.
-        
-        The mesh includes:
-        - vertices: 3D coordinates of each vertex
-        - faces: triangles defined by vertex indices
-        - face_colors: RGBA colors for each face based on sim_grid values
-        - visual: trimesh.visual.ColorVisuals object storing the face colors
-    
-    Examples
-    --------
-    Basic usage with Green View Index data:
-    >>> gvi = np.array([[0.5, 0.6], [0.4, 0.8]])  # GVI values
-    >>> dem = np.array([[10.0, 10.2], [9.8, 10.1]])  # Ground heights
-    >>> mesh = create_sim_surface_mesh(gvi, dem, meshsize=1.0, z_offset=1.5)
-    
-    Custom color range and colormap:
-    >>> mesh = create_sim_surface_mesh(gvi, dem,
-    ...                               cmap_name='RdYlBu',
-    ...                               vmin=0.0, vmax=1.0)
-    
-    Notes
-    -----
-    - The function automatically creates a matplotlib colorbar figure for visualization
-    - Both input grids are flipped vertically to match the voxel_array orientation
-    - Each grid cell is converted to two triangles for compatibility with 3D engines
-    - The mesh is positioned at dem_grid + z_offset to float above the terrain
-    - Face colors are interpolated from the colormap based on sim_grid values
-    """
-    # Flip arrays vertically
-    sim_grid_flipped = np.flipud(sim_grid)
-    dem_grid_flipped = np.flipud(dem_grid)
-
-    # Identify valid (non-NaN) values
-    valid_mask = ~np.isnan(sim_grid_flipped)
-    valid_values = sim_grid_flipped[valid_mask]
-    if valid_values.size == 0:
-        return None
-
-    # If vmin/vmax not provided, use actual min/max of the valid sim data
-    if vmin is None:
-        vmin = np.nanmin(valid_values) 
-    if vmax is None:
-        vmax = np.nanmax(valid_values)
-        
-    # Prepare the colormap and create colorbar
-    norm = mcolors.Normalize(vmin=vmin, vmax=vmax)
-    scalar_map = cm.ScalarMappable(norm=norm, cmap=cmap_name)
-    
-    # Create a figure just for the colorbar
-    fig, ax = plt.subplots(figsize=(6, 1))
-    plt.colorbar(scalar_map, cax=ax, orientation='horizontal')
-    plt.tight_layout()
-    plt.close()
-
-    vertices = []
-    faces = []
-    face_colors = []
-
-    vert_index = 0
-    nrows, ncols = sim_grid_flipped.shape
-
-    # Build a quad (two triangles) for each valid cell
-    for x in range(nrows):
-        for y in range(ncols):
-            val = sim_grid_flipped[x, y]
-            if np.isnan(val):
-                continue
-
-            z_base = meshsize * int(dem_grid_flipped[x, y] / meshsize + 1.5) + z_offset            
-
-            # 4 corners in (x,y)*meshsize
-            v0 = [ x      * meshsize,  y      * meshsize, z_base ]
-            v1 = [(x + 1) * meshsize,  y      * meshsize, z_base ]
-            v2 = [(x + 1) * meshsize, (y + 1) * meshsize, z_base ]
-            v3 = [ x      * meshsize, (y + 1) * meshsize, z_base ]
-
-            vertices.extend([v0, v1, v2, v3])
-            faces.extend([
-                [vert_index,     vert_index + 1, vert_index + 2],
-                [vert_index,     vert_index + 2, vert_index + 3]
-            ])
-
-            # Get color from colormap
-            color_rgba = np.array(scalar_map.to_rgba(val))  # shape (4,)
-
-            # Each cell has 2 faces => add the color twice
-            face_colors.append(color_rgba)
-            face_colors.append(color_rgba)
-
-            vert_index += 4
-
-    if len(vertices) == 0:
-        return None
-
-    vertices = np.array(vertices, dtype=float)
-    faces = np.array(faces, dtype=int)
-    face_colors = np.array(face_colors, dtype=float)
-
-    mesh = trimesh.Trimesh(
-        vertices=vertices,
-        faces=faces,
-        face_colors=face_colors,
-        process=False  # skip auto merge if you want to preserve quads
-    )
-
-    return mesh
-
-def create_city_meshes(
-    voxel_array,
-    vox_dict,
-    meshsize=1.0,
-    include_classes=None,
-    exclude_classes=None,
-):
-    """
-    Create a collection of colored 3D meshes representing different city elements.
-    
-    This function processes a voxelized city model and creates separate meshes for
-    different urban elements (buildings, trees, etc.), each with its own color.
-    The function preserves sharp edges and applies appropriate colors based on the
-    provided color dictionary.
-    
-    Parameters
-    ----------
-    voxel_array : np.ndarray (3D)
-        3D array representing the voxelized city model. Each voxel contains a class ID
-        that maps to an urban element type:
-        - 0: Void/air (automatically skipped)
-        - -2: Trees
-        - -3: Buildings
-        Other values can represent different urban elements as defined in vox_dict.
-    
-    vox_dict : dict
-        Dictionary mapping class IDs to RGB colors. Each entry should be:
-        {class_id: [R, G, B]} where R, G, B are 0-255 integer values.
-        Example: {-3: [200, 200, 200], -2: [0, 255, 0]} for grey buildings and
-        green trees. The key 0 (air) is automatically excluded.
-    
-    meshsize : float, default=1.0
-        Size of each voxel in meters, applied uniformly to x, y, and z dimensions.
-        Used to scale the output meshes to real-world coordinates.
-    
-    Returns
-    -------
-    meshes : dict
-        Dictionary mapping class IDs to their corresponding trimesh.Trimesh objects.
-        Each mesh includes:
-        - vertices: 3D coordinates scaled by meshsize
-        - faces: triangulated faces preserving sharp edges
-        - face_colors: RGBA colors from vox_dict
-        - visual: trimesh.visual.ColorVisuals object storing the face colors
-        
-        Classes with no voxels are automatically excluded from the output.
-    
-    Examples
-    --------
-    Basic usage with buildings and trees:
-    >>> voxels = np.zeros((10, 10, 10))
-    >>> voxels[4:7, 4:7, 0:5] = -3  # Add a building
-    >>> voxels[2:4, 2:4, 0:3] = -2  # Add some trees
-    >>> colors = {
-    ...     -3: [200, 200, 200],  # Grey buildings
-    ...     -2: [0, 255, 0]       # Green trees
-    ... }
-    >>> meshes = create_city_meshes(voxels, colors, meshsize=1.0)
-    
-    Notes
-    -----
-    - The function automatically skips class_id=0 (typically air/void)
-    - Each urban element type gets its own separate mesh for efficient rendering
-    - Colors are converted from RGB [0-255] to RGBA [0-1] format
-    - Sharp edges are preserved to maintain architectural features
-    - Empty classes (no voxels) are automatically excluded from the output
-    - Errors during mesh creation for a class are caught and reported
-    """
-    meshes = {}
-
-    # Convert RGB colors to hex for material properties
-    color_dict = {k: mcolors.rgb2hex([v[0]/255, v[1]/255, v[2]/255])
-                 for k, v in vox_dict.items() if k != 0}  # Exclude air
-
-    # Determine which classes to process
-    unique_classes = np.unique(voxel_array)
-
-    if include_classes is not None:
-        # Only keep classes explicitly requested (and present in the data)
-        class_iterable = [c for c in include_classes if c in unique_classes]
-    else:
-        class_iterable = list(unique_classes)
-
-    exclude_set = set(exclude_classes) if exclude_classes is not None else set()
-
-    # Create vertices and faces for each object class
-    for class_id in class_iterable:
-        if class_id == 0:  # Skip air
-            continue
-
-        if class_id in exclude_set:
-            # Explicitly skipped (e.g., will be replaced with custom mesh)
-            continue
-
-        try:
-            mesh = create_voxel_mesh(voxel_array, class_id, meshsize=meshsize)
-
-            if mesh is None:
-                continue
-
-            # Convert hex color to RGBA
-            if class_id not in color_dict:
-                # Color not provided; skip silently for robustness
-                continue
-            rgb_color = np.array(mcolors.hex2color(color_dict[class_id]))
-            rgba_color = np.concatenate([rgb_color, [1.0]])
-
-            # Assign color to all faces
-            mesh.visual.face_colors = np.tile(rgba_color, (len(mesh.faces), 1))
-
-            meshes[class_id] = mesh
-
-        except ValueError as e:
-            print(f"Skipping class {class_id}: {e}")
-
-    return meshes
-
-def export_meshes(meshes, output_directory, base_filename):
-    """
-    Export a collection of meshes to both OBJ (with MTL) and STL formats.
-    
-    This function exports meshes in two ways:
-    1. A single combined OBJ file with materials (and associated MTL file)
-    2. Separate STL files for each mesh, named with their class IDs
-    
-    Parameters
-    ----------
-    meshes : dict
-        Dictionary mapping class IDs to trimesh.Trimesh objects.
-        Each mesh should have:
-        - vertices: 3D coordinates
-        - faces: triangulated faces
-        - face_colors: RGBA colors (if using materials)
-    
-    output_directory : str
-        Directory path where the output files will be saved.
-        Will be created if it doesn't exist.
-    
-    base_filename : str
-        Base name for the output files (without extension).
-        Will be used to create:
-        - {base_filename}.obj : Combined mesh with materials
-        - {base_filename}.mtl : Material definitions for OBJ
-        - {base_filename}_{class_id}.stl : Individual STL files
-    
-    Returns
-    -------
-    None
-        Files are written directly to the specified output directory.
-    
-    Examples
-    --------
-    >>> meshes = {
-    ...     -3: building_mesh,  # Building mesh with grey color
-    ...     -2: tree_mesh      # Tree mesh with green color
-    ... }
-    >>> export_meshes(meshes, 'output/models', 'city_model')
-    
-    This will create:
-    - output/models/city_model.obj
-    - output/models/city_model.mtl
-    - output/models/city_model_-3.stl
-    - output/models/city_model_-2.stl
-    
-    Notes
-    -----
-    - OBJ/MTL format preserves colors and materials but is more complex
-    - STL format is simpler but doesn't support colors
-    - STL files are exported separately for each class for easier processing
-    - The OBJ file combines all meshes while preserving their materials
-    - File extensions are automatically added to the base filename
-    """
-    # Export combined mesh as OBJ with materials
-    combined_mesh = trimesh.util.concatenate(list(meshes.values()))
-    combined_mesh.export(f"{output_directory}/{base_filename}.obj")
-
-    # Export individual meshes as STL
-    for class_id, mesh in meshes.items():
-        # Convert class_id to a string for filename
-        mesh.export(f"{output_directory}/{base_filename}_{class_id}.stl")
-
-def split_vertices_manual(mesh):
-    """
-    Split a mesh into independent faces by duplicating shared vertices.
-    
-    This function imitates trimesh's split_vertices() functionality but ensures
-    complete face independence by giving each face its own copy of vertices.
-    This is particularly useful for rendering applications where smooth shading
-    between faces is undesirable, such as architectural visualization in Rhino.
-    
-    Parameters
-    ----------
-    mesh : trimesh.Trimesh
-        Input mesh to split. Should have:
-        - vertices: array of vertex coordinates
-        - faces: array of vertex indices forming triangles
-        - visual: Optional ColorVisuals object with face colors
-    
-    Returns
-    -------
-    out_mesh : trimesh.Trimesh
-        New mesh where each face is completely independent, with:
-        - Duplicated vertices for each face
-        - No vertex sharing between faces
-        - Preserved face colors if present in input
-        - Each face as a separate component
-    
-    Examples
-    --------
-    Basic usage:
-    >>> vertices = np.array([[0,0,0], [1,0,0], [1,1,0], [0,1,0]])
-    >>> faces = np.array([[0,1,2], [0,2,3]])  # Two triangles sharing vertices
-    >>> mesh = trimesh.Trimesh(vertices=vertices, faces=faces)
-    >>> split_mesh = split_vertices_manual(mesh)
-    >>> print(f"Original vertices: {len(mesh.vertices)}")  # 4 vertices
-    >>> print(f"Split vertices: {len(split_mesh.vertices)}")  # 6 vertices
-    
-    With face colors:
-    >>> colors = np.array([[255,0,0,255], [0,255,0,255]])  # Red and green faces
-    >>> mesh.visual = trimesh.visual.ColorVisuals(mesh, face_colors=colors)
-    >>> split_mesh = split_vertices_manual(mesh)  # Colors are preserved
-    
-    Notes
-    -----
-    - Each output face has exactly 3 unique vertices
-    - Face colors are preserved in the output mesh
-    - Useful for:
-        - Preventing smooth shading artifacts
-        - Ensuring face color independence
-        - Preparing meshes for CAD software
-        - Creating sharp edges in architectural models
-    - Memory usage increases as vertices are duplicated
-    """
-    new_meshes = []
-    
-    # For each face, build a small, one-face mesh
-    for face_idx, face in enumerate(mesh.faces):
-        face_coords = mesh.vertices[face]
-        
-        # Create mini-mesh without colors first
-        mini_mesh = trimesh.Trimesh(
-            vertices=face_coords,
-            faces=[[0, 1, 2]],
-            process=False  # skip merging/cleaning
-        )
-        
-        # If the mesh has per-face colors, set the face color properly
-        if (mesh.visual.face_colors is not None 
-            and len(mesh.visual.face_colors) == len(mesh.faces)):
-            # Create a visual object with the face color (for one face)
-            face_color = mesh.visual.face_colors[face_idx]
-            color_visual = trimesh.visual.ColorVisuals(
-                mesh=mini_mesh,
-                face_colors=np.array([face_color]),  # One face, one color
-                vertex_colors=None
-            )
-            mini_mesh.visual = color_visual
-        
-        new_meshes.append(mini_mesh)
-    
-    # Concatenate all the single-face meshes
-    out_mesh = trimesh.util.concatenate(new_meshes)
-    return out_mesh
-
-def save_obj_from_colored_mesh(meshes, output_path, base_filename, max_materials=None):
-    """
-    Memory-safe OBJ/MTL exporter.
-    - Streams vertices/faces to disk (no concatenate, no per-face mini-meshes).
-    - Uses face colors -> materials (no vertex splitting).
-    - Optional color quantization to reduce material count.
-    """
-    import os
-    import numpy as np
-
-    os.makedirs(output_path, exist_ok=True)
-    obj_path = os.path.join(output_path, f"{base_filename}.obj")
-    mtl_path = os.path.join(output_path, f"{base_filename}.mtl")
-
-    # --------------- helpers ---------------
-    def to_uint8_rgba(arr):
-        arr = np.asarray(arr)
-        if arr.dtype != np.uint8:
-            # Handle float [0..1] or int [0..255]
-            if arr.dtype.kind == 'f':
-                arr = np.clip(arr, 0.0, 1.0)
-                arr = (arr * 255.0 + 0.5).astype(np.uint8)
-            else:
-                arr = arr.astype(np.uint8)
-        if arr.shape[1] == 3:
-            alpha = np.full((arr.shape[0], 1), 255, dtype=np.uint8)
-            arr = np.concatenate([arr, alpha], axis=1)
-        return arr
-
-    # First pass: build material palette
-    # We avoid collecting all colors at once—scan per mesh and update a dict.
-    color_to_id = {}
-    ordered_colors = []  # list of RGBA uint8 tuples in material order
-
-    # Optional quantizer (lazy-init)
-    quantizer = None
-    if max_materials is not None:
-        try:
-            from sklearn.cluster import MiniBatchKMeans
-            quantizer = MiniBatchKMeans(n_clusters=max_materials, random_state=42, batch_size=8192)
-            # Partial-fit streaming pass over colors
-            for m in meshes.values():
-                fc = getattr(m.visual, "face_colors", None)
-                if fc is None:
-                    continue
-                fc = to_uint8_rgba(fc)
-                if fc.size == 0: 
-                    continue
-                # Use only RGB for clustering
-                quantizer.partial_fit(fc[:, :3].astype(np.float32))
-        except ImportError:
-            raise ImportError("scikit-learn is required for color quantization. Install it with: pip install scikit-learn")
-
-    # Assign material ids during a second scan, but still streaming to avoid big unions
-    def get_material_id(rgba):
-        key = (int(rgba[0]), int(rgba[1]), int(rgba[2]), int(rgba[3]))
-        mid = color_to_id.get(key)
-        if mid is None:
-            mid = len(ordered_colors)
-            color_to_id[key] = mid
-            ordered_colors.append(key)
-        return mid
-
-    # 2nd pass if quantizing: we need color centroids
-    centers_u8 = None
-    if quantizer is not None:
-        centers = quantizer.cluster_centers_.astype(np.float32)  # RGB float
-        centers = np.clip(centers, 0.0, 255.0).astype(np.uint8)
-        # Build a quick LUT fun
-        def quantize_rgb(rgb_u8):
-            # rgb_u8: (N,3) uint8 -> labels -> centers
-            labels = quantizer.predict(rgb_u8.astype(np.float32))
-            return centers[labels]
-        # We'll convert each mesh's face colors to quantized RGB on the fly
-        centers_u8 = centers
-
-    # Build materials palette by scanning once (still O(total faces) but tiny memory)
-    for m in meshes.values():
-        fc = getattr(m.visual, "face_colors", None)
-        if fc is None:
-            # No colors: assign default grey
-            rgba = np.array([[200,200,200,255]], dtype=np.uint8)
-            get_material_id(rgba[0])
-            continue
-        fc = to_uint8_rgba(fc)
-        if quantizer is not None:
-            q_rgb = quantize_rgb(fc[:, :3])
-            fc = np.concatenate([q_rgb, fc[:, 3:4]], axis=1)
-        # Iterate unique colors in this mesh to limit get_material_id calls
-        # but don't materialize huge sets; unique per mesh is fine.
-        uniq = np.unique(fc, axis=0)
-        for rgba in uniq:
-            get_material_id(rgba)
-
-    # Write MTL
-    with open(mtl_path, "w") as mtl:
-        for i, (r, g, b, a) in enumerate(ordered_colors):
-            mtl.write(f"newmtl material_{i}\n")
-            # Match viewport look: diffuse only, no specular. Many viewers assume sRGB.
-            kd_r, kd_g, kd_b = r/255.0, g/255.0, b/255.0
-            mtl.write(f"Kd {kd_r:.6f} {kd_g:.6f} {kd_b:.6f}\n")
-            # Ambient same as diffuse to avoid darkening in some viewers
-            mtl.write(f"Ka {kd_r:.6f} {kd_g:.6f} {kd_b:.6f}\n")
-            # No specular highlight
-            mtl.write("Ks 0.000000 0.000000 0.000000\n")
-            # Disable lighting model with specular; keep simple shading
-            mtl.write("illum 1\n")
-            # Alpha
-            mtl.write(f"d {a/255.0:.6f}\n\n")
-
-    # Stream OBJ
-    with open(obj_path, "w") as obj:
-        obj.write(f"mtllib {os.path.basename(mtl_path)}\n")
-
-        v_offset = 0  # running vertex index offset
-        # Reusable cache so we don't keep writing 'usemtl' for the same block unnecessarily
-        current_material = None
-
-        for class_id, m in meshes.items():
-            verts = np.asarray(m.vertices, dtype=np.float64)
-            faces = np.asarray(m.faces, dtype=np.int64)
-            if verts.size == 0 or faces.size == 0:
-                continue
-
-            # Write vertices
-            # (We do a single pass; writing text is the bottleneck, but memory-safe.)
-            for v in verts:
-                obj.write(f"v {v[0]:.6f} {v[1]:.6f} {v[2]:.6f}\n")
-
-            # Prepare face colors (face-level)
-            fc = getattr(m.visual, "face_colors", None)
-            if fc is None or len(fc) != len(faces):
-                # default grey if missing or mismatched
-                fc = np.tile(np.array([200,200,200,255], dtype=np.uint8), (len(faces), 1))
-            else:
-                fc = to_uint8_rgba(fc)
-
-            if quantizer is not None:
-                q_rgb = quantize_rgb(fc[:, :3])
-                fc = np.concatenate([q_rgb, fc[:, 3:4]], axis=1)
-
-            # Group faces by material id and stream in order
-            # Build material id per face quickly
-            # Convert face colors to material ids
-            # (Avoid Python loops over faces more than once)
-            # Map unique colors in this mesh to material ids first:
-            uniq_colors, inv_idx = np.unique(fc, axis=0, return_inverse=True)
-            color_to_mid_local = {tuple(c.tolist()): get_material_id(c) for c in uniq_colors}
-            mids = np.fromiter(
-                (color_to_mid_local[tuple(c.tolist())] for c in uniq_colors[inv_idx]),
-                dtype=np.int64,
-                count=len(inv_idx)
-            )
-
-            # Write faces grouped by material, but preserve simple ordering
-            # Cheap approach: emit runs; switching material only when necessary
-            current_material = None
-            for i_face, face in enumerate(faces):
-                mid = int(mids[i_face])
-                if current_material != mid:
-                    obj.write(f"usemtl material_{mid}\n")
-                    current_material = mid
-                a, b, c = face + 1 + v_offset  # OBJ is 1-based
-                obj.write(f"f {a} {b} {c}\n")
-
-            v_offset += len(verts)
-
+import numpy as np
+import os
+import trimesh
+import matplotlib.colors as mcolors
+import matplotlib.cm as cm
+import matplotlib.pyplot as plt
+
+def create_voxel_mesh(voxel_array, class_id, meshsize=1.0, building_id_grid=None, mesh_type=None):
+    """
+    Create a 3D mesh from voxels preserving sharp edges, scaled by meshsize.
+    
+    This function converts a 3D voxel array into a triangulated mesh, where each voxel
+    face is converted into two triangles. The function preserves sharp edges between
+    different classes and handles special cases for buildings.
+    
+    Parameters
+    ----------
+    voxel_array : np.ndarray (3D)
+        The voxel array of shape (X, Y, Z) where each cell contains a class ID.
+        - 0: typically represents void/air
+        - -2: typically represents trees
+        - -3: typically represents buildings
+        Other values can represent different classes as defined by the application.
+    
+    class_id : int
+        The ID of the class to extract. Only voxels with this ID will be included
+        in the output mesh.
+    
+    meshsize : float, default=1.0
+        The real-world size of each voxel in meters, applied uniformly to x, y, and z
+        dimensions. Used to scale the output mesh to real-world coordinates.
+    
+    building_id_grid : np.ndarray (2D), optional
+        2D grid of building IDs with shape (X, Y). Only used when class_id=-3 (buildings).
+        Each cell contains a unique identifier for the building at that location.
+        This allows tracking which faces belong to which building.
+    
+    mesh_type : str, optional
+        Type of mesh to create, controlling which faces are included:
+        - None (default): create faces at boundaries between different classes
+        - 'building_solar': only create faces at boundaries between buildings (-3)
+                          and either void (0) or trees (-2). Useful for solar analysis
+                          where only exposed surfaces matter.
+
+    Returns
+    -------
+    mesh : trimesh.Trimesh or None
+        The resulting triangulated mesh for the given class_id. Returns None if no
+        voxels of the specified class are found.
+        
+        The mesh includes:
+        - vertices: 3D coordinates of each vertex
+        - faces: triangles defined by vertex indices
+        - face_normals: normal vectors for each face
+        - metadata: If class_id=-3, includes 'building_id' mapping faces to buildings
+    
+    Examples
+    --------
+    Basic usage for a simple voxel array:
+    >>> voxels = np.zeros((10, 10, 10))
+    >>> voxels[4:7, 4:7, 0:5] = 1  # Create a simple column
+    >>> mesh = create_voxel_mesh(voxels, class_id=1, meshsize=0.5)
+    
+    Creating a building mesh with IDs:
+    >>> building_ids = np.zeros((10, 10))
+    >>> building_ids[4:7, 4:7] = 1  # Mark building #1
+    >>> mesh = create_voxel_mesh(voxels, class_id=-3, 
+    ...                         building_id_grid=building_ids,
+    ...                         meshsize=1.0)
+    
+    Notes
+    -----
+    - The function creates faces only at boundaries between different classes or at
+      the edges of the voxel array.
+    - Each face is split into two triangles for compatibility with graphics engines.
+    - Face normals are computed to ensure correct lighting and rendering.
+    - For buildings (class_id=-3), building IDs are tracked to maintain building identity.
+    - The mesh preserves sharp edges, which is important for architectural visualization.
+    """
+    # Find voxels of the current class
+    voxel_coords = np.argwhere(voxel_array == class_id)
+
+    if building_id_grid is not None:
+        building_id_grid_flipud = np.flipud(building_id_grid)
+
+    if len(voxel_coords) == 0:
+        return None
+
+    # Define the 6 faces of a unit cube (local coordinates 0..1)
+    unit_faces = np.array([
+        # Front
+        [[0, 0, 1], [1, 0, 1], [1, 1, 1], [0, 1, 1]],
+        # Back
+        [[0, 0, 0], [0, 1, 0], [1, 1, 0], [1, 0, 0]],
+        # Right
+        [[1, 0, 0], [1, 1, 0], [1, 1, 1], [1, 0, 1]],
+        # Left
+        [[0, 0, 0], [0, 0, 1], [0, 1, 1], [0, 1, 0]],
+        # Top
+        [[0, 1, 0], [0, 1, 1], [1, 1, 1], [1, 1, 0]],
+        # Bottom
+        [[0, 0, 0], [1, 0, 0], [1, 0, 1], [0, 0, 1]]
+    ])
+
+    # Define face normals
+    face_normals = np.array([
+        [0, 0, 1],   # Front
+        [0, 0, -1],  # Back
+        [1, 0, 0],   # Right
+        [-1, 0, 0],  # Left
+        [0, 1, 0],   # Top
+        [0, -1, 0]   # Bottom
+    ])
+
+    vertices = []
+    faces = []
+    face_normals_list = []
+    building_ids = []  # List to store building IDs for each face
+
+    for x, y, z in voxel_coords:
+        # For buildings, get the building ID from the grid
+        building_id = None
+        if class_id == -3 and building_id_grid is not None:
+            building_id = building_id_grid_flipud[x, y]
+            
+        # Check each face of the current voxel
+        adjacent_coords = [
+            (x,   y,   z+1),  # Front
+            (x,   y,   z-1),  # Back
+            (x+1, y,   z),    # Right
+            (x-1, y,   z),    # Left
+            (x,   y+1, z),    # Top
+            (x,   y-1, z)     # Bottom
+        ]
+
+        # Only create faces where there's a transition based on mesh_type
+        for face_idx, adj_coord in enumerate(adjacent_coords):
+            try:
+                # If adj_coord is outside array bounds, it's a boundary => face is visible
+                if adj_coord[0] < 0 or adj_coord[1] < 0 or adj_coord[2] < 0:
+                    is_boundary = True
+                else:
+                    adj_value = voxel_array[adj_coord]
+                    
+                    if mesh_type == 'open_air' and class_id == -3:
+                        # For building_solar, only create faces at boundaries with void (0) or trees (-2)
+                        is_boundary = (adj_value == 0 or adj_value == -2)
+                    else:
+                        # Default behavior - create faces at any class change
+                        is_boundary = (adj_value == 0 or adj_value != class_id)
+            except IndexError:
+                # Out of range => boundary
+                is_boundary = True
+
+            if is_boundary:
+                # Local face in (0..1) for x,y,z, then shift by voxel coords
+                face_verts = (unit_faces[face_idx] + np.array([x, y, z])) * meshsize
+                current_vert_count = len(vertices)
+
+                vertices.extend(face_verts)
+                # Convert quad to two triangles
+                faces.extend([
+                    [current_vert_count, current_vert_count + 1, current_vert_count + 2],
+                    [current_vert_count, current_vert_count + 2, current_vert_count + 3]
+                ])
+                # Add face normals for both triangles
+                face_normals_list.extend([face_normals[face_idx], face_normals[face_idx]])
+                
+                # Store building ID for both triangles if this is a building
+                if class_id == -3 and building_id_grid is not None:
+                    building_ids.extend([building_id, building_id])
+
+    if not vertices:
+        return None
+
+    vertices = np.array(vertices)
+    faces = np.array(faces)
+    face_normals_list = np.array(face_normals_list)
+
+    # Create mesh
+    mesh = trimesh.Trimesh(
+        vertices=vertices,
+        faces=faces,
+        face_normals=face_normals_list
+    )
+
+    # Merge vertices that are at the same position
+    mesh.merge_vertices()
+    
+    # Ensure metadata dict exists
+    if not hasattr(mesh, 'metadata') or mesh.metadata is None:
+        mesh.metadata = {}
+
+    # Store intended per-triangle normals to avoid reliance on auto-computed normals
+    mesh.metadata['provided_face_normals'] = face_normals_list
+
+    # Add building IDs as metadata for buildings
+    if class_id == -3 and building_id_grid is not None and building_ids:
+        mesh.metadata['building_id'] = np.array(building_ids)
+
+    return mesh
+
+def create_sim_surface_mesh(sim_grid, dem_grid,
+                            meshsize=1.0, z_offset=1.5,
+                            cmap_name='viridis',
+                            vmin=None, vmax=None):
+    """
+    Create a colored planar surface mesh from simulation data, positioned above a DEM.
+    
+    This function generates a 3D visualization mesh for 2D simulation results (like
+    Green View Index, solar radiation, etc.). The mesh is positioned above the Digital
+    Elevation Model (DEM) by a specified offset, and colored according to the simulation
+    values using a matplotlib colormap.
+    
+    Parameters
+    ----------
+    sim_grid : 2D np.ndarray
+        2D array of simulation values (e.g., Green View Index, solar radiation).
+        NaN values in this grid will be skipped in the output mesh.
+        The grid should be oriented with north at the top.
+    
+    dem_grid : 2D np.ndarray
+        2D array of ground elevations in meters. Must have the same shape as sim_grid.
+        Used to position the visualization mesh at the correct height above terrain.
+    
+    meshsize : float, default=1.0
+        Size of each cell in meters. Applied uniformly to x and y dimensions.
+        Determines the resolution of the output mesh.
+    
+    z_offset : float, default=1.5
+        Additional height offset in meters added to dem_grid elevations.
+        Used to position the visualization above ground level for better visibility.
+    
+    cmap_name : str, default='viridis'
+        Matplotlib colormap name used for coloring the mesh based on sim_grid values.
+        Common options:
+        - 'viridis': Default, perceptually uniform, colorblind-friendly
+        - 'RdYlBu': Red-Yellow-Blue, good for diverging data
+        - 'jet': Rainbow colormap (not recommended for scientific visualization)
+    
+    vmin : float, optional
+        Minimum value for color mapping. If None, uses min of sim_grid (excluding NaN).
+        Used to control the range of the colormap.
+    
+    vmax : float, optional
+        Maximum value for color mapping. If None, uses max of sim_grid (excluding NaN).
+        Used to control the range of the colormap.
+    
+    Returns
+    -------
+    mesh : trimesh.Trimesh or None
+        A single mesh containing one colored square face (two triangles) per non-NaN cell.
+        Returns None if there are no valid (non-NaN) cells in sim_grid.
+        
+        The mesh includes:
+        - vertices: 3D coordinates of each vertex
+        - faces: triangles defined by vertex indices
+        - face_colors: RGBA colors for each face based on sim_grid values
+        - visual: trimesh.visual.ColorVisuals object storing the face colors
+    
+    Examples
+    --------
+    Basic usage with Green View Index data:
+    >>> gvi = np.array([[0.5, 0.6], [0.4, 0.8]])  # GVI values
+    >>> dem = np.array([[10.0, 10.2], [9.8, 10.1]])  # Ground heights
+    >>> mesh = create_sim_surface_mesh(gvi, dem, meshsize=1.0, z_offset=1.5)
+    
+    Custom color range and colormap:
+    >>> mesh = create_sim_surface_mesh(gvi, dem,
+    ...                               cmap_name='RdYlBu',
+    ...                               vmin=0.0, vmax=1.0)
+    
+    Notes
+    -----
+    - The function automatically creates a matplotlib colorbar figure for visualization
+    - Both input grids are flipped vertically to match the voxel_array orientation
+    - Each grid cell is converted to two triangles for compatibility with 3D engines
+    - The mesh is positioned at dem_grid + z_offset to float above the terrain
+    - Face colors are interpolated from the colormap based on sim_grid values
+    """
+    # Flip arrays vertically
+    sim_grid_flipped = np.flipud(sim_grid)
+    dem_grid_flipped = np.flipud(dem_grid)
+
+    # Identify valid (non-NaN) values
+    valid_mask = ~np.isnan(sim_grid_flipped)
+    valid_values = sim_grid_flipped[valid_mask]
+    if valid_values.size == 0:
+        return None
+
+    # If vmin/vmax not provided, use actual min/max of the valid sim data
+    if vmin is None:
+        vmin = np.nanmin(valid_values) 
+    if vmax is None:
+        vmax = np.nanmax(valid_values)
+        
+    # Prepare the colormap and create colorbar
+    norm = mcolors.Normalize(vmin=vmin, vmax=vmax)
+    scalar_map = cm.ScalarMappable(norm=norm, cmap=cmap_name)
+    
+    # Create a figure just for the colorbar
+    fig, ax = plt.subplots(figsize=(6, 1))
+    plt.colorbar(scalar_map, cax=ax, orientation='horizontal')
+    plt.tight_layout()
+    plt.close()
+
+    vertices = []
+    faces = []
+    face_colors = []
+
+    vert_index = 0
+    nrows, ncols = sim_grid_flipped.shape
+
+    # Build a quad (two triangles) for each valid cell
+    for x in range(nrows):
+        for y in range(ncols):
+            val = sim_grid_flipped[x, y]
+            if np.isnan(val):
+                continue
+
+            z_base = meshsize * int(dem_grid_flipped[x, y] / meshsize + 1.5) + z_offset            
+
+            # 4 corners in (x,y)*meshsize
+            v0 = [ x      * meshsize,  y      * meshsize, z_base ]
+            v1 = [(x + 1) * meshsize,  y      * meshsize, z_base ]
+            v2 = [(x + 1) * meshsize, (y + 1) * meshsize, z_base ]
+            v3 = [ x      * meshsize, (y + 1) * meshsize, z_base ]
+
+            vertices.extend([v0, v1, v2, v3])
+            faces.extend([
+                [vert_index,     vert_index + 1, vert_index + 2],
+                [vert_index,     vert_index + 2, vert_index + 3]
+            ])
+
+            # Get color from colormap
+            color_rgba = np.array(scalar_map.to_rgba(val))  # shape (4,)
+
+            # Each cell has 2 faces => add the color twice
+            face_colors.append(color_rgba)
+            face_colors.append(color_rgba)
+
+            vert_index += 4
+
+    if len(vertices) == 0:
+        return None
+
+    vertices = np.array(vertices, dtype=float)
+    faces = np.array(faces, dtype=int)
+    face_colors = np.array(face_colors, dtype=float)
+
+    mesh = trimesh.Trimesh(
+        vertices=vertices,
+        faces=faces,
+        face_colors=face_colors,
+        process=False  # skip auto merge if you want to preserve quads
+    )
+
+    return mesh
+
+def create_city_meshes(
+    voxel_array,
+    vox_dict,
+    meshsize=1.0,
+    include_classes=None,
+    exclude_classes=None,
+):
+    """
+    Create a collection of colored 3D meshes representing different city elements.
+    
+    This function processes a voxelized city model and creates separate meshes for
+    different urban elements (buildings, trees, etc.), each with its own color.
+    The function preserves sharp edges and applies appropriate colors based on the
+    provided color dictionary.
+    
+    Parameters
+    ----------
+    voxel_array : np.ndarray (3D)
+        3D array representing the voxelized city model. Each voxel contains a class ID
+        that maps to an urban element type:
+        - 0: Void/air (automatically skipped)
+        - -2: Trees
+        - -3: Buildings
+        Other values can represent different urban elements as defined in vox_dict.
+    
+    vox_dict : dict
+        Dictionary mapping class IDs to RGB colors. Each entry should be:
+        {class_id: [R, G, B]} where R, G, B are 0-255 integer values.
+        Example: {-3: [200, 200, 200], -2: [0, 255, 0]} for grey buildings and
+        green trees. The key 0 (air) is automatically excluded.
+    
+    meshsize : float, default=1.0
+        Size of each voxel in meters, applied uniformly to x, y, and z dimensions.
+        Used to scale the output meshes to real-world coordinates.
+    
+    Returns
+    -------
+    meshes : dict
+        Dictionary mapping class IDs to their corresponding trimesh.Trimesh objects.
+        Each mesh includes:
+        - vertices: 3D coordinates scaled by meshsize
+        - faces: triangulated faces preserving sharp edges
+        - face_colors: RGBA colors from vox_dict
+        - visual: trimesh.visual.ColorVisuals object storing the face colors
+        
+        Classes with no voxels are automatically excluded from the output.
+    
+    Examples
+    --------
+    Basic usage with buildings and trees:
+    >>> voxels = np.zeros((10, 10, 10))
+    >>> voxels[4:7, 4:7, 0:5] = -3  # Add a building
+    >>> voxels[2:4, 2:4, 0:3] = -2  # Add some trees
+    >>> colors = {
+    ...     -3: [200, 200, 200],  # Grey buildings
+    ...     -2: [0, 255, 0]       # Green trees
+    ... }
+    >>> meshes = create_city_meshes(voxels, colors, meshsize=1.0)
+    
+    Notes
+    -----
+    - The function automatically skips class_id=0 (typically air/void)
+    - Each urban element type gets its own separate mesh for efficient rendering
+    - Colors are converted from RGB [0-255] to RGBA [0-1] format
+    - Sharp edges are preserved to maintain architectural features
+    - Empty classes (no voxels) are automatically excluded from the output
+    - Errors during mesh creation for a class are caught and reported
+    """
+    meshes = {}
+
+    # Convert RGB colors to hex for material properties
+    color_dict = {k: mcolors.rgb2hex([v[0]/255, v[1]/255, v[2]/255])
+                 for k, v in vox_dict.items() if k != 0}  # Exclude air
+
+    # Determine which classes to process
+    unique_classes = np.unique(voxel_array)
+
+    if include_classes is not None:
+        # Only keep classes explicitly requested (and present in the data)
+        class_iterable = [c for c in include_classes if c in unique_classes]
+    else:
+        class_iterable = list(unique_classes)
+
+    exclude_set = set(exclude_classes) if exclude_classes is not None else set()
+
+    # Create vertices and faces for each object class
+    for class_id in class_iterable:
+        if class_id == 0:  # Skip air
+            continue
+
+        if class_id in exclude_set:
+            # Explicitly skipped (e.g., will be replaced with custom mesh)
+            continue
+
+        try:
+            mesh = create_voxel_mesh(voxel_array, class_id, meshsize=meshsize)
+
+            if mesh is None:
+                continue
+
+            # Convert hex color to RGBA
+            if class_id not in color_dict:
+                # Color not provided; skip silently for robustness
+                continue
+            rgb_color = np.array(mcolors.hex2color(color_dict[class_id]))
+            rgba_color = np.concatenate([rgb_color, [1.0]])
+
+            # Assign color to all faces
+            mesh.visual.face_colors = np.tile(rgba_color, (len(mesh.faces), 1))
+
+            meshes[class_id] = mesh
+
+        except ValueError as e:
+            print(f"Skipping class {class_id}: {e}")
+
+    return meshes
+
+def export_meshes(meshes, output_directory, base_filename):
+    """
+    Export a collection of meshes to both OBJ (with MTL) and STL formats.
+    
+    This function exports meshes in two ways:
+    1. A single combined OBJ file with materials (and associated MTL file)
+    2. Separate STL files for each mesh, named with their class IDs
+    
+    Parameters
+    ----------
+    meshes : dict
+        Dictionary mapping class IDs to trimesh.Trimesh objects.
+        Each mesh should have:
+        - vertices: 3D coordinates
+        - faces: triangulated faces
+        - face_colors: RGBA colors (if using materials)
+    
+    output_directory : str
+        Directory path where the output files will be saved.
+        Will be created if it doesn't exist.
+    
+    base_filename : str
+        Base name for the output files (without extension).
+        Will be used to create:
+        - {base_filename}.obj : Combined mesh with materials
+        - {base_filename}.mtl : Material definitions for OBJ
+        - {base_filename}_{class_id}.stl : Individual STL files
+    
+    Returns
+    -------
+    None
+        Files are written directly to the specified output directory.
+    
+    Examples
+    --------
+    >>> meshes = {
+    ...     -3: building_mesh,  # Building mesh with grey color
+    ...     -2: tree_mesh      # Tree mesh with green color
+    ... }
+    >>> export_meshes(meshes, 'output/models', 'city_model')
+    
+    This will create:
+    - output/models/city_model.obj
+    - output/models/city_model.mtl
+    - output/models/city_model_-3.stl
+    - output/models/city_model_-2.stl
+    
+    Notes
+    -----
+    - OBJ/MTL format preserves colors and materials but is more complex
+    - STL format is simpler but doesn't support colors
+    - STL files are exported separately for each class for easier processing
+    - The OBJ file combines all meshes while preserving their materials
+    - File extensions are automatically added to the base filename
+    """
+    # Export combined mesh as OBJ with materials
+    combined_mesh = trimesh.util.concatenate(list(meshes.values()))
+    combined_mesh.export(f"{output_directory}/{base_filename}.obj")
+
+    # Export individual meshes as STL
+    for class_id, mesh in meshes.items():
+        # Convert class_id to a string for filename
+        mesh.export(f"{output_directory}/{base_filename}_{class_id}.stl")
+
+def split_vertices_manual(mesh):
+    """
+    Split a mesh into independent faces by duplicating shared vertices.
+    
+    This function imitates trimesh's split_vertices() functionality but ensures
+    complete face independence by giving each face its own copy of vertices.
+    This is particularly useful for rendering applications where smooth shading
+    between faces is undesirable, such as architectural visualization in Rhino.
+    
+    Parameters
+    ----------
+    mesh : trimesh.Trimesh
+        Input mesh to split. Should have:
+        - vertices: array of vertex coordinates
+        - faces: array of vertex indices forming triangles
+        - visual: Optional ColorVisuals object with face colors
+    
+    Returns
+    -------
+    out_mesh : trimesh.Trimesh
+        New mesh where each face is completely independent, with:
+        - Duplicated vertices for each face
+        - No vertex sharing between faces
+        - Preserved face colors if present in input
+        - Each face as a separate component
+    
+    Examples
+    --------
+    Basic usage:
+    >>> vertices = np.array([[0,0,0], [1,0,0], [1,1,0], [0,1,0]])
+    >>> faces = np.array([[0,1,2], [0,2,3]])  # Two triangles sharing vertices
+    >>> mesh = trimesh.Trimesh(vertices=vertices, faces=faces)
+    >>> split_mesh = split_vertices_manual(mesh)
+    >>> print(f"Original vertices: {len(mesh.vertices)}")  # 4 vertices
+    >>> print(f"Split vertices: {len(split_mesh.vertices)}")  # 6 vertices
+    
+    With face colors:
+    >>> colors = np.array([[255,0,0,255], [0,255,0,255]])  # Red and green faces
+    >>> mesh.visual = trimesh.visual.ColorVisuals(mesh, face_colors=colors)
+    >>> split_mesh = split_vertices_manual(mesh)  # Colors are preserved
+    
+    Notes
+    -----
+    - Each output face has exactly 3 unique vertices
+    - Face colors are preserved in the output mesh
+    - Useful for:
+        - Preventing smooth shading artifacts
+        - Ensuring face color independence
+        - Preparing meshes for CAD software
+        - Creating sharp edges in architectural models
+    - Memory usage increases as vertices are duplicated
+    """
+    new_meshes = []
+    
+    # For each face, build a small, one-face mesh
+    for face_idx, face in enumerate(mesh.faces):
+        face_coords = mesh.vertices[face]
+        
+        # Create mini-mesh without colors first
+        mini_mesh = trimesh.Trimesh(
+            vertices=face_coords,
+            faces=[[0, 1, 2]],
+            process=False  # skip merging/cleaning
+        )
+        
+        # If the mesh has per-face colors, set the face color properly
+        if (mesh.visual.face_colors is not None 
+            and len(mesh.visual.face_colors) == len(mesh.faces)):
+            # Create a visual object with the face color (for one face)
+            face_color = mesh.visual.face_colors[face_idx]
+            color_visual = trimesh.visual.ColorVisuals(
+                mesh=mini_mesh,
+                face_colors=np.array([face_color]),  # One face, one color
+                vertex_colors=None
+            )
+            mini_mesh.visual = color_visual
+        
+        new_meshes.append(mini_mesh)
+    
+    # Concatenate all the single-face meshes
+    out_mesh = trimesh.util.concatenate(new_meshes)
+    return out_mesh
+
+def save_obj_from_colored_mesh(meshes, output_path, base_filename, max_materials=None):
+    """
+    Memory-safe OBJ/MTL exporter.
+    - Streams vertices/faces to disk (no concatenate, no per-face mini-meshes).
+    - Uses face colors -> materials (no vertex splitting).
+    - Optional color quantization to reduce material count.
+    """
+    import os
+    import numpy as np
+
+    os.makedirs(output_path, exist_ok=True)
+    obj_path = os.path.join(output_path, f"{base_filename}.obj")
+    mtl_path = os.path.join(output_path, f"{base_filename}.mtl")
+
+    # --------------- helpers ---------------
+    def to_uint8_rgba(arr):
+        arr = np.asarray(arr)
+        if arr.dtype != np.uint8:
+            # Handle float [0..1] or int [0..255]
+            if arr.dtype.kind == 'f':
+                arr = np.clip(arr, 0.0, 1.0)
+                arr = (arr * 255.0 + 0.5).astype(np.uint8)
+            else:
+                arr = arr.astype(np.uint8)
+        if arr.shape[1] == 3:
+            alpha = np.full((arr.shape[0], 1), 255, dtype=np.uint8)
+            arr = np.concatenate([arr, alpha], axis=1)
+        return arr
+
+    # First pass: build material palette
+    # We avoid collecting all colors at once—scan per mesh and update a dict.
+    color_to_id = {}
+    ordered_colors = []  # list of RGBA uint8 tuples in material order
+
+    # Optional quantizer (lazy-init)
+    quantizer = None
+    if max_materials is not None:
+        try:
+            from sklearn.cluster import MiniBatchKMeans
+            quantizer = MiniBatchKMeans(n_clusters=max_materials, random_state=42, batch_size=8192)
+            # Partial-fit streaming pass over colors
+            for m in meshes.values():
+                fc = getattr(m.visual, "face_colors", None)
+                if fc is None:
+                    continue
+                fc = to_uint8_rgba(fc)
+                if fc.size == 0: 
+                    continue
+                # Use only RGB for clustering
+                quantizer.partial_fit(fc[:, :3].astype(np.float32))
+        except ImportError:
+            raise ImportError("scikit-learn is required for color quantization. Install it with: pip install scikit-learn")
+
+    # Assign material ids during a second scan, but still streaming to avoid big unions
+    def get_material_id(rgba):
+        key = (int(rgba[0]), int(rgba[1]), int(rgba[2]), int(rgba[3]))
+        mid = color_to_id.get(key)
+        if mid is None:
+            mid = len(ordered_colors)
+            color_to_id[key] = mid
+            ordered_colors.append(key)
+        return mid
+
+    # 2nd pass if quantizing: we need color centroids
+    centers_u8 = None
+    if quantizer is not None:
+        centers = quantizer.cluster_centers_.astype(np.float32)  # RGB float
+        centers = np.clip(centers, 0.0, 255.0).astype(np.uint8)
+        # Build a quick LUT fun
+        def quantize_rgb(rgb_u8):
+            # rgb_u8: (N,3) uint8 -> labels -> centers
+            labels = quantizer.predict(rgb_u8.astype(np.float32))
+            return centers[labels]
+        # We'll convert each mesh's face colors to quantized RGB on the fly
+        centers_u8 = centers
+
+    # Build materials palette by scanning once (still O(total faces) but tiny memory)
+    for m in meshes.values():
+        fc = getattr(m.visual, "face_colors", None)
+        if fc is None:
+            # No colors: assign default grey
+            rgba = np.array([[200,200,200,255]], dtype=np.uint8)
+            get_material_id(rgba[0])
+            continue
+        fc = to_uint8_rgba(fc)
+        if quantizer is not None:
+            q_rgb = quantize_rgb(fc[:, :3])
+            fc = np.concatenate([q_rgb, fc[:, 3:4]], axis=1)
+        # Iterate unique colors in this mesh to limit get_material_id calls
+        # but don't materialize huge sets; unique per mesh is fine.
+        uniq = np.unique(fc, axis=0)
+        for rgba in uniq:
+            get_material_id(rgba)
+
+    # Write MTL
+    with open(mtl_path, "w") as mtl:
+        for i, (r, g, b, a) in enumerate(ordered_colors):
+            mtl.write(f"newmtl material_{i}\n")
+            # Match viewport look: diffuse only, no specular. Many viewers assume sRGB.
+            kd_r, kd_g, kd_b = r/255.0, g/255.0, b/255.0
+            mtl.write(f"Kd {kd_r:.6f} {kd_g:.6f} {kd_b:.6f}\n")
+            # Ambient same as diffuse to avoid darkening in some viewers
+            mtl.write(f"Ka {kd_r:.6f} {kd_g:.6f} {kd_b:.6f}\n")
+            # No specular highlight
+            mtl.write("Ks 0.000000 0.000000 0.000000\n")
+            # Disable lighting model with specular; keep simple shading
+            mtl.write("illum 1\n")
+            # Alpha
+            mtl.write(f"d {a/255.0:.6f}\n\n")
+
+    # Stream OBJ
+    with open(obj_path, "w") as obj:
+        obj.write(f"mtllib {os.path.basename(mtl_path)}\n")
+
+        v_offset = 0  # running vertex index offset
+        # Reusable cache so we don't keep writing 'usemtl' for the same block unnecessarily
+        current_material = None
+
+        for class_id, m in meshes.items():
+            verts = np.asarray(m.vertices, dtype=np.float64)
+            faces = np.asarray(m.faces, dtype=np.int64)
+            if verts.size == 0 or faces.size == 0:
+                continue
+
+            # Write vertices
+            # (We do a single pass; writing text is the bottleneck, but memory-safe.)
+            for v in verts:
+                obj.write(f"v {v[0]:.6f} {v[1]:.6f} {v[2]:.6f}\n")
+
+            # Prepare face colors (face-level)
+            fc = getattr(m.visual, "face_colors", None)
+            if fc is None or len(fc) != len(faces):
+                # default grey if missing or mismatched
+                fc = np.tile(np.array([200,200,200,255], dtype=np.uint8), (len(faces), 1))
+            else:
+                fc = to_uint8_rgba(fc)
+
+            if quantizer is not None:
+                q_rgb = quantize_rgb(fc[:, :3])
+                fc = np.concatenate([q_rgb, fc[:, 3:4]], axis=1)
+
+            # Group faces by material id and stream in order
+            # Build material id per face quickly
+            # Convert face colors to material ids
+            # (Avoid Python loops over faces more than once)
+            # Map unique colors in this mesh to material ids first:
+            uniq_colors, inv_idx = np.unique(fc, axis=0, return_inverse=True)
+            color_to_mid_local = {tuple(c.tolist()): get_material_id(c) for c in uniq_colors}
+            mids = np.fromiter(
+                (color_to_mid_local[tuple(c.tolist())] for c in uniq_colors[inv_idx]),
+                dtype=np.int64,
+                count=len(inv_idx)
+            )
+
+            # Write faces grouped by material, but preserve simple ordering
+            # Cheap approach: emit runs; switching material only when necessary
+            current_material = None
+            for i_face, face in enumerate(faces):
+                mid = int(mids[i_face])
+                if current_material != mid:
+                    obj.write(f"usemtl material_{mid}\n")
+                    current_material = mid
+                a, b, c = face + 1 + v_offset  # OBJ is 1-based
+                obj.write(f"f {a} {b} {c}\n")
+
+            v_offset += len(verts)
+
     return obj_path, mtl_path