PointNet++

monroyaume

已于 2022-04-03 07:25:48 修改

阅读量1.1k

点赞数 1

分类专栏：机器学习入门文章标签：检测

于 2022-04-03 06:29:10 首次发布

本文链接：https://blog.csdn.net/king52113141314/article/details/123932185

版权

机器学习入门专栏收录该内容

45 篇文章

订阅专栏

1,furthest point sampling or farthest point sampling

Furthest Point Sampling | Minibatch AI

def fps(points, n_samples):
    """
    points: [N, 3] array containing the whole point cloud
    n_samples: samples you want in the sampled point cloud typically << N 
    """
    points = np.array(points)
    
    # Represent the points by their indices in points
    points_left = np.arange(len(points)) # [P]

    # Initialise an array for the sampled indices
    sample_inds = np.zeros(n_samples, dtype='int') # [S]

    # Initialise distances to inf
    dists = np.ones_like(points_left) * float('inf') # [P]

    # Select a point from points by its index, save it
    selected = 0
    sample_inds[0] = points_left[selected]

    # Delete selected 
    points_left = np.delete(points_left, selected) # [P - 1]

    # Iteratively select points for a maximum of n_samples
    for i in range(1, n_samples):
        # Find the distance to the last added point in selected
        # and all the others
        last_added = sample_inds[i-1]
        
        dist_to_last_added_point = (
            (points[last_added] - points[points_left])**2).sum(-1) # [P - i]

        # If closer, updated distances
        dists[points_left] = np.minimum(dist_to_last_added_point, 
                                        dists[points_left]) # [P - i]

        # We want to pick the one that has the largest nearest neighbour
        # distance to the sampled points
        selected = np.argmax(dists[points_left])
        sample_inds[i] = points_left[selected]

        # Update points_left
        points_left = np.delete(points_left, selected)

    return points[sample_inds]