3D点云深度学习算法入门指南：从数据预处理到模型训练实战

1次阅读

共计 3402 个字符，预计需要花费 9 分钟才能阅读完成。

3D 点云是由激光雷达或深度相机采集的离散三维坐标集合，具有以下核心特征：

无序性(Unordered): 点云是坐标的集合而非序列，排列顺序不影响几何含义。这意味着算法需要对排列顺序具有不变性。
非结构化(Non-grid): 不同于图像像素的规则排列，点云在空间中呈不规则分布。
旋转不变性(Rotation Invariance): 同一物体的点云在不同视角下应被识别为相同类别。

这些特性给传统深度学习带来了两大挑战：
1. 常规卷积神经网络 (CNN) 无法直接处理无序点集
2. 需要设计具有几何变换不变性的特征提取器

算法名称	核心思想	适用场景	优缺点
PointNet	对称函数 + 全局特征池化	分类 / 简单分割	计算高效但局部特征提取弱
PointNet++	层次化特征学习 +FPS 采样	复杂分割 / 细粒度分类	多尺度特征但计算成本较高
PointCNN	χ-Conv 学习排列不变性变换	密集点云处理	内存占用大实现复杂
DGCNN	动态图卷积构建局部邻域	语义分割	对噪声敏感

import open3d as o3d
import numpy as np

def normalize_point_cloud(points):
    """
    点云标准化处理
    输入: [N, 3] numpy 数组
    输出: 归一化到单位球体的点云
    """
    centroid = np.mean(points, axis=0)
    points -= centroid
    max_dist = np.max(np.sqrt(np.sum(points**2, axis=1)))
    points /= max_dist
    return points

def fps_sampling(points, n_samples):
    """
    最远点采样(Farthest Point Sampling)
    输入: [N,3]点云, 采样点数 K
    输出: [K,3]采样点索引
    """
    sampled_indices = np.zeros(n_samples, dtype=np.int32)
    distances = np.full(points.shape[0], np.inf)
    # 随机选择初始点
    farthest = np.random.randint(0, points.shape[0])
    for i in range(n_samples):
        sampled_indices[i] = farthest
        dist = np.sum((points - points[farthest])**2, axis=1)
        mask = dist < distances
        distances[mask] = dist[mask]
        farthest = np.argmax(distances)
    return sampled_indices

import torch
import torch.nn as nn

class PointNetSetAbstraction(nn.Module):
    """
    层级特征提取模块
    实现采样 (Sampling)、分组(Grouping) 和特征提取
    """
    def __init__(self, n_point, radius, n_sample, mlp):
        super().__init__()
        self.n_point = n_point
        self.radius = radius
        self.n_sample = n_sample
        self.mlp_convs = nn.ModuleList()
        last_channel = 3  # xyz 初始维度
        for out_channel in mlp:
            self.mlp_convs.append(nn.Conv2d(last_channel, out_channel, 1))
            self.mlp_convs.append(nn.BatchNorm2d(out_channel))
            self.mlp_convs.append(nn.ReLU())
            last_channel = out_channel

    def forward(self, xyz, points):
        """
        输入: 
            xyz: [B, N, 3] 点云坐标
            points: [B, N, C] 点特征
        输出:
            new_xyz: [B, S, 3] 采样点坐标
            new_points: [B, S, C'] 聚合特征"""
        B, N, C = xyz.shape
        S = self.n_point
        # 最远点采样
        fps_idx = farthest_point_sample(xyz, S)  # [B,S]
        new_xyz = index_points(xyz, fps_idx)     # [B,S,3]
        # 球查询构造局部区域
        idx = query_ball_point(self.radius, self.n_sample, xyz, new_xyz)
        grouped_xyz = index_points(xyz, idx)     # [B,S,nsample,3]
        grouped_xyz -= new_xyz.unsqueeze(2)      # 相对坐标
        if points is not None:
            grouped_points = index_points(points, idx)
            grouped_points = torch.cat([grouped_points, grouped_xyz], dim=-1)
        else:
            grouped_points = grouped_xyz
        # 特征提取 MLP
        grouped_points = grouped_points.permute(0,3,1,2)  # [B,C,S,nsample]
        for conv in self.mlp_convs:
            grouped_points = conv(grouped_points)
        # 最大池化获得局部特征
        new_points = torch.max(grouped_points, 3)[0]  # [B,C,S]
        return new_xyz, new_points.permute(0,2,1)

def random_rotate_point_cloud(points):
    """随机旋转增强"""
    rotation_angle = np.random.uniform() * 2 * np.pi
    cosval = np.cos(rotation_angle)
    sinval = np.sin(rotation_angle)
    rotation_matrix = np.array([[cosval, sinval, 0],
        [-sinval, cosval, 0],
        [0, 0, 1]])
    return np.dot(points, rotation_matrix)

def jitter_point_cloud(points, sigma=0.01, clip=0.05):
    """添加随机噪声"""
    N, C = points.shape
    noise = np.clip(sigma * np.random.randn(N, C), -clip, clip)
    return points + noise