修改参数

2026-11-12 15:32:02 +08:00
1 parent 2434c4140d
commit 46bf872bac
@@ -0,0 +1,155 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""
+fill_gap_v3.py — 基于射线投射的眼镜-脸空隙填充
+"""
+
+import numpy as np
+from scipy.spatial import cKDTree
+from scipy.ndimage import binary_dilation, distance_transform_edt
+from skimage import measure
+import trimesh
+from trimesh.smoothing import filter_taubin
+import argparse
+import time
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--obj", required=True)
+    parser.add_argument("--index", required=True)
+    parser.add_argument("--out", default="fill_mesh.obj")
+    parser.add_argument("--res", type=int, default=60, help="分辨率（先跑低分辨率验证）")
+    parser.add_argument("--gap", type=float, default=1.5, help="填充间隙宽度 mm")
+    args = parser.parse_args()
+
+    t0 = time.time()
+    print("=== 眼镜-脸部空隙填充 v3 ===\n")
+
+    # 1. 加载
+    mesh = trimesh.load(args.obj, force="mesh")
+    V = mesh.vertices
+    F = mesh.faces
+    idx = np.loadtxt(args.index, dtype=np.int64) - 1
+    print(f"[load] 顶点 {len(V)}, 面 {len(F)}, 眼镜顶点 {len(idx)}")
+
+    # 2. 分类
+    is_glass = np.zeros(len(V), dtype=bool)
+    is_glass[idx] = True
+    glass_verts = V[idx]
+    face_verts = V[~is_glass]
+    print(f"[classify] 眼镜顶点 {len(glass_verts)}, 脸部顶点 {len(face_verts)}")
+
+    # 3. 对每个眼镜顶点，找最近的脸部顶点
+    face_tree = cKDTree(face_verts)
+    dists, nearest_idx = face_tree.query(glass_verts, k=1)
+    nearest_face_verts = face_verts[nearest_idx]
+    
+    # 方向向量：从脸部指向眼镜
+    directions = glass_verts - nearest_face_verts
+    dir_lengths = np.linalg.norm(directions, axis=1)
+    directions = directions / dir_lengths[:, None]
+    
+    # 统计平均方向
+    avg_dir = directions.mean(axis=0)
+    avg_dir = avg_dir / np.linalg.norm(avg_dir)
+    print(f"[dir] 平均方向（脸部→眼镜）: {avg_dir.round(3)}")
+
+    # 4. 确定填充范围
+    # 眼镜中心
+    glass_center = glass_verts.mean(axis=0)
+    # 沿方向投影
+    proj = (glass_verts - glass_center) @ avg_dir
+    front_pos = proj.max()  # 眼镜前表面（远离脸部）
+    back_pos = proj.min()   # 眼镜后表面（靠近脸部）
+    print(f"[surf] 前 {front_pos:.2f}, 后 {back_pos:.2f}")
+
+    # 5. 包围盒（只用眼镜区域）
+    margin = 4.0
+    lo = glass_verts.min(axis=0) - margin
+    hi = glass_verts.max(axis=0) + margin
+    print(f"[bbox] {lo.round(2)} -> {hi.round(2)}")
+
+    # 6. 体素网格（低分辨率测试）
+    shape = np.array([args.res] * 3)
+    pitch = (hi - lo).max() / args.res
+    xs = np.linspace(lo[0], hi[0], shape[0])
+    ys = np.linspace(lo[1], hi[1], shape[1])
+    zs = np.linspace(lo[2], hi[2], shape[2])
+    X, Y, Z = np.meshgrid(xs, ys, zs, indexing="ij")
+    points = np.stack([X.ravel(), Y.ravel(), Z.ravel()], axis=1)
+    print(f"[grid] {shape} = {points.shape[0]} 体素, 大小 {pitch:.3f}mm")
+
+    # 7. KD-Tree
+    glass_tree = cKDTree(glass_verts)
+    face_tree = cKDTree(face_verts)
+    print("[tree] KD-Tree 构建完成")
+
+    # 8. 计算距离
+    print("[dist] 计算距离场...")
+    dist_to_glass, _ = glass_tree.query(points, k=1)
+    dist_to_face, _ = face_tree.query(points, k=1)
+
+    # 9. 计算每个体素在"方向"上的投影
+    proj_val = (points - glass_center) @ avg_dir
+
+    # 10. 填充规则
+    # 在眼镜后表面附近（靠近脸部的那侧）
+    near_glass = dist_to_glass < args.gap * 2
+    
+    # 不在脸内部
+    outside_face = dist_to_face > 0.3
+    
+    # 在后表面附近（back_pos 到 back_pos + gap 之间）
+    in_gap = (proj_val > back_pos - 0.5) & (proj_val < back_pos + args.gap)
+    
+    # 组合
+    fill_grid = near_glass & outside_face & in_gap
+    fill_grid = fill_grid.reshape(shape)
+
+    n_fill = fill_grid.sum()
+    print(f"[fill] 填充体素: {n_fill}")
+
+    if n_fill < 30:
+        print("[warn] 太少，尝试仅用近眼镜条件...")
+        fill_grid = near_glass.reshape(shape)
+        n_fill = fill_grid.sum()
+        print(f"[fill] 仅近眼镜: {n_fill}")
+
+    if n_fill < 10:
+        print("[error] 填充为空")
+        return
+
+    # 11. 膨胀
+    fill_grid = binary_dilation(fill_grid, iterations=3)
+
+    # 12. MC
+    print("[mc] Marching Cubes...")
+    sdf = distance_transform_edt(~fill_grid).astype(np.float32)
+    sdf[fill_grid] = -distance_transform_edt(fill_grid)[fill_grid]
+
+    verts, faces, _, _ = measure.marching_cubes(
+        sdf, level=0.0,
+        spacing=(pitch, pitch, pitch)
+    )
+    verts += lo
+
+    result = trimesh.Trimesh(vertices=verts, faces=faces, process=True)
+
+    # 取最大分量
+    if len(result.split()) > 1:
+        comps = result.split(only_watertight=False)
+        result = max(comps, key=lambda m: m.volume)
+        print(f"[clean] 取最大分量")
+
+    filter_taubin(result, lamb=0.5, nu=0.5, iterations=5)
+
+    result.export(args.out)
+    print(f"\n[output] {args.out}")
+    print(f"  顶点: {len(result.vertices)}, 面: {len(result.faces)}")
+    print(f"  体积: {result.volume:.3f}, 水密: {result.is_watertight}")
+    print(f"  耗时: {time.time()-t0:.1f}s")
+
+
+if __name__ == "__main__":
+    main()