代码学习_KPConv_models.blocks

最新推荐文章于 2024-10-01 16:24:17 发布

Kougami..

最新推荐文章于 2024-10-01 16:24:17 发布

阅读量904

点赞数 3

文章标签：深度学习

本文链接：https://blog.csdn.net/weixin_44109545/article/details/115605354

版权

本文是作者关于深度学习中KPConv模型的学习笔记，主要关注点在于KPConv模块的初始化、邻居点的选择及其特征计算过程。介绍了如何从点云数据中选取邻域点，计算点与核的距离，应用核函数进行特征加权，以及最终的特征归一化。文中还提及了邻居点的选取方法，但表示这一部分有待进一步学习。

摘要由CSDN通过智能技术生成

文章只是个人学习过程中的记录，因水平极其有限，仅供参考，欢迎一起交流

class KPConv(nn.Module)

初始化赋值
创建kernel points: K_points_numpy

        # Initialize weights  [n_kpoints, in_fdim, out_fdim]
        self.weights = Parameter(torch.zeros((self.K, in_channels, out_channels), dtype=torch.float32),
                                 requires_grad=True)
                                 
		# Initialize kernel points  [num_kpoints, dimension=3]
        self.kernel_points = Parameter(torch.tensor(K_points_numpy, dtype=torch.float32),
                         requires_grad=False)

对于点云中某点x，取出其邻域内的点，并与自身作差 ~ y_i
- 细节见附1
计算其与 kernel 的距离的平方 ~ sq_distances

        # Get neighbor points
        neighbors = s_pts[neighb_inds, :]  # [n_points, n_neighbors, dim]

        # Center every neighborhood
        neighbors = neighbors - q_pts.unsqueeze_(1)  # [n_points, n_neighbors, dim]

        # kernel_points [num_kpoints, dim=3]
        deformed_K_points = self.kernel_points # [num_kpoints, dim]

        # Get all difference matrices [n_points, n_neighbors, n_kpoints, dim]
        neighbors.unsqueeze_(2)
        differences = neighbors - deformed_K_points

        # Get the square distances [n_points, n_neighbors, n_kpoints]
        sq_distances = torch.sum