欣赏高手的逻辑。
如何计算点云中每个A和任一个其他点B,和 B和其紧邻点B^ 连接向量 的夹角呢?
知识点包括 :
1. 通过向量叉乘计算向量夹角正弦Sin
2. 通过向量点乘计算向量夹角余弦Cos
3. 通过上述两者计算夹角。
不得不佩服ZhengQin的代码能力,确实有点东西
k = self.angle_k
knn_indices = dist_map.topk(k=k + 1, dim=2, largest=False)[1][:, :, 1:] # (B, N, k)
knn_indices = knn_indices.unsqueeze(3).expand(batch_size, num_point, k, 3) # (B, N, k, 3)
expanded_points = points.unsqueeze(1).expand(batch_size, num_point, num_point, 3) # (B, N, N, 3)
knn_points = torch.gather(expanded_points, dim=2, index=knn_indices) # (B, N, k, 3)
ref_vectors = knn_points - points.unsqueeze(2) # (B, N, k, 3)
anc_vectors = points.unsqueeze(1) - points.unsqueeze(2) # (B, N, N, 3)
# 只是为了计算,拓展为想同维度。
ref_vectors = ref_vectors.unsqueeze(2).expand(batch_size, num_point, num_point, k, 3) # (B, N, N, k, 3)
anc_vectors = anc_vectors.unsqueeze(3).expand(batch_size, num_point, num_point, k, 3) # (B, N, N, k, 3)
# 向量乘积来计算角度。
sin_values = torch.linalg.norm(torch.cross(ref_vectors, anc_vectors, dim=-1), dim=-1) # (B, N, N, k)
cos_values = torch.sum(ref_vectors * anc_vectors, dim=-1) # (B, N, N, k)
angles = torch.atan2(sin_values, cos_values) # (B, N, N, k)