ScanNetpp-SpatialLM / align_scannetpp.py
Valentina-Zhang's picture
Upload align_scannetpp.py
d82ae7a verified
import open3d as o3d
import numpy as np
import os
def align_scene_to_z_up(pcd, save_pcd_path=None, save_transform_path=None, visualize=False, fit_ground=True):
"""
将点云扫描转换为Z轴向上对齐,并尝试将墙壁对齐到X-Y平面
参数:
points (np.ndarray): (N, 3) 点云数据
visualize (bool): 是否可视化结果
返回:
aligned_points (np.ndarray): 对齐后的点云
transform_matrix (np.ndarray): 应用的4x4变换矩阵
"""
points = np.asarray(pcd.points)
centroid = np.mean(points, axis=0)
distance_threshold = 0.02 # 动态距离阈值
R_ground = np.eye(3) # Initialize ground rotation matrix
if fit_ground:
plane_model, ground_inliers = pcd.segment_plane(
distance_threshold=distance_threshold,
ransac_n=3,
num_iterations=1000
)
[a, b, c, d] = plane_model
ground_normal = np.array([a, b, c])
# print('ground_normal:', ground_normal)
# 可视化地面点云
ground_cloud = pcd.select_by_index(ground_inliers)
ground_cloud.paint_uniform_color([1, 0, 0]) # 红色表示地面
if visualize:
o3d.visualization.draw_geometries([ground_cloud])
# 3. 计算旋转矩阵使地面法向量对齐到Z轴
z_axis = np.array([0, 0, 1])
ground_normal = ground_normal / np.linalg.norm(ground_normal)
# 计算旋转轴和角度
rotation_axis = np.cross(ground_normal, z_axis)
if np.linalg.norm(rotation_axis) < 1e-6:
rotation_axis = np.array([0, 1, 0]) # 避免零向量
else:
rotation_axis = rotation_axis / np.linalg.norm(rotation_axis)
cos_theta = np.dot(ground_normal, z_axis)
angle = np.arccos(np.clip(cos_theta, -1.0, 1.0))
# 使用罗德里格斯公式计算旋转矩阵
K = np.array([
[0, -rotation_axis[2], rotation_axis[1]],
[rotation_axis[2], 0, -rotation_axis[0]],
[-rotation_axis[1], rotation_axis[0], 0]
])
R_ground = np.eye(3) + np.sin(angle) * K + (1 - np.cos(angle)) * (K @ K)
# [Alternate] concise method
# from scipy.spatial.transform import Rotation as R
# R_ground = R.from_rotvec(rotation_axis * angle).as_matrix()
# 4. 应用地面旋转
centered_points = points - centroid
ground_rotated = (R_ground @ centered_points.T).T
# 5. 检测墙壁平面(垂直平面)
all_indices = np.arange(len(points))
non_ground_indices = np.setdiff1d(all_indices, ground_inliers)
non_ground_points = ground_rotated[non_ground_indices]
# 创建临时点云用于墙壁检测
temp_pcd = o3d.geometry.PointCloud()
temp_pcd.points = o3d.utility.Vector3dVector(non_ground_points)
remaining_pcd = temp_pcd
else:
# already z up
remaining_pcd = pcd
wall_planes = []
wall_directions = []
max_points = 0
best_direction = None
R_walls = np.eye(3)
# 检测多个墙壁平面
for _ in range(6):
if len(remaining_pcd.points) < 100:
break
plane_model, inliers = remaining_pcd.segment_plane(
distance_threshold=distance_threshold,
ransac_n=3,
num_iterations=1000
)
wall_cloud = remaining_pcd.select_by_index(inliers)
wall_cloud.paint_uniform_color([1, 0, 0]) # 红色表示地面
if visualize:
o3d.visualization.draw_geometries([wall_cloud])
[a, b, c, d] = plane_model
normal = np.array([a, b, c])
# 检查是否为垂直平面(法向量的Z分量接近0)
if abs(normal[2]) < 0.1 and np.linalg.norm(normal[:2]) > 0.5:
# 提取水平方向
horizontal_dir = normal[:2] / np.linalg.norm(normal[:2])
wall_directions = horizontal_dir
# print('wall direction',normal)
if max_points < len(inliers):
max_points = len(inliers)
# print(max_points)
best_direction = wall_directions
# 计算最佳方向的旋转角度
angle = np.arctan2(best_direction[1], best_direction[0])
# 创建最终的旋转矩阵
R_walls = np.array([
[np.cos(-angle), -np.sin(-angle), 0],
[np.sin(-angle), np.cos(-angle), 0],
[0, 0, 1]
])
remaining_pcd = remaining_pcd.select_by_index(inliers, invert=True)
# 8. 创建变换矩阵(旋转 + 平移)
if fit_ground:
# Compose the transformations: first ground rotation, then wall rotation
combined_rotation = R_walls @ R_ground
transform_matrix = np.eye(4)
transform_matrix[:3, :3] = combined_rotation
transform_matrix[:3, 3] = -combined_rotation @ centroid
else:
# Only wall rotation
transform_matrix = np.eye(4)
transform_matrix[:3, :3] = R_walls
transform_matrix[:3, 3] = -R_walls @ centroid # 平移使中心到原点
# 9. 应用变换
aligned_points = (transform_matrix[:3, :3] @ points.T + transform_matrix[:3, 3:4]).T
# 10. 创建对齐后的点云
aligned_pcd = o3d.geometry.PointCloud()
aligned_pcd.points = o3d.utility.Vector3dVector(aligned_points)
if pcd.colors:
aligned_pcd.colors = pcd.colors
# 11. 保存结果
if save_pcd_path:
o3d.io.write_point_cloud(save_pcd_path, aligned_pcd)
if save_transform_path:
np.savetxt(save_transform_path, transform_matrix, fmt='%.8f')
return aligned_pcd, transform_matrix
"""创建4x4变换矩阵"""
transform = np.eye(4)
transform[:3, :3] = rotation_matrix
return transform
def visualize_alignment(pcd_orig, pcd_aligned):
"""可视化原始点云和对齐后的点云"""
pcd_orig.paint_uniform_color([1, 0, 0]) # 红色为原始点云
pcd_aligned.paint_uniform_color([0, 1, 0]) # 绿色为对齐后点云
# 创建坐标系
coord_frame = o3d.geometry.TriangleMesh.create_coordinate_frame(size=1.0)
o3d.visualization.draw_geometries([pcd_orig, pcd_aligned, coord_frame])
npy_dir = '/media/vivo/vivo/Datasets/ScanNetpp/ScanNetpp_preprocessed/train/'
scene_names = os.listdir(npy_dir)
for scene_name in scene_names:
if scene_name.endswith('.zip'):
continue
print(scene_name)
scene_dir = os.path.join(npy_dir, scene_name)
npy_path = os.path.join(scene_dir, 'coord.npy')
points = np.load(npy_path) # (N, 3) 或 (N, 6)
# 创建点云对象
pcd = o3d.geometry.PointCloud()
pcd.points = o3d.utility.Vector3dVector(points[:, :3])
if points.shape[1] == 6:
pcd.colors = o3d.utility.Vector3dVector(points[:, 3:6]) # 如果包含 RGB
# 执行对齐
transformed_pcd, transform = align_scene_to_z_up(
pcd,
# save_transform_path=os.path.join(scene_dir, "transform.txt"),
visualize=False,
fit_ground=False
)
aligned_points = np.asarray(transformed_pcd.points)
# scale point cloud
# min_z = np.min(aligned_points[:, 2])
# max_z = np.max(aligned_points[:, 2])
# height = max_z - min_z
# print(f"Original height: {height}")
# estimated_height = 2.5
# scale = estimated_height / height
# print(f"Scale factor: {scale}")
# aligned_points_scaled = aligned_points * scale
transformed_pcd.points = o3d.utility.Vector3dVector(aligned_points)
# 保存为对齐后的 PLY 文件
# aligned_ply_path = os.path.join(scene_dir, scene_name + ".ply")
# o3d.io.write_point_cloud(aligned_ply_path, transformed_pcd)
# 也可以保存为对齐后的 npy
np.save(os.path.join(scene_dir, "coord_align.npy"), aligned_points)