点云最远点采样（Python）

最远点采样（FPS）可以设置采样的点数，比起网格采样等方法更加实用，被应用在点云处理方法中（例如PointNet++）。缺点是每次要计算两个集合中所有点的相对距离，计算量较大。但实际上，采取矩阵运算的实现方式，计算速率也能接受。

下面是FPS的Python实现，主要包括计算索引和映射点两步骤，度量使用欧式距离。

def FarthestPointSampling_ForBatch(xyz, npoint):
    B, N, C = xyz.shape

    centroids = np.zeros((B, npoint))
    distance = np.ones((B, N)) * 1e10

    batch_indices = np.arange(B)

    barycenter = np.sum((xyz), 1)
    barycenter = barycenter/xyz.shape[1]
    barycenter = barycenter.reshape(B, 1, 3)

    dist = np.sum((xyz - barycenter) ** 2, -1)
    farthest = np.argmax(dist,1)

    for i in range(npoint):
        print("-------------------------------------------------------")
        print("The %d farthest pts %s " % (i, farthest))
        centroids[:, i] = farthest
        centroid = xyz[batch_indices, farthest, :].reshape(B, 1, 3)
        dist = np.sum((xyz - centroid) ** 2, -1)
        mask = dist < distance
        distance[mask] = dist[mask]
        farthest = np.argmax(distance, -1)
    return centroids.astype('int')

def index_points(points, idx):
    batch_indices = np.arange(points.shape[0])
    batch_indices = batch_indices.repeat(idx.shape[1]).reshape((-1,idx.shape[1]))  #(batch_size,npoints)
    print(batch_indices.shape)
    new_points = points[batch_indices,idx, :]
    return new_points

参考资料