r"""Sinkhorn Optimal transport with dustbin parameter (SuperGlue style).

r"""Point Matching with Local-to-Global Registration. Args: k (int): top-k selection for point matching. acceptance_r

Initialize parameters for KPConv. Modified from [KPConv-PyTorch](https://github.com/HuguesTHOMAS/KPConv-PyTorch). Deformable KPConv

r"""Initialize a standard unary block with GroupNorm and LeakyReLU. Args: in_channels: dimension input features out_c

r"""Initialize a standard last_unary block without GN, ReLU. Args: in_channels: dimension input features out_channels

r"""Initialize a KPConv block with ReLU and BatchNorm. Args: in_channels: dimension input features out_channels: dime

r"""Initialize a ResNet bottleneck block. Args: in_channels: dimension input features out_channels: dimension input f

* Called during search to add an element matching the criteria. * @return true if the search should be continued, false if the results are * suf

Add points to the set, Inserts all points from [start, end] */

r"""Average reduce all tensors across all workers.

r"""Rotate points and normals (optional) along the origin. Given a point cloud P(3, N), normals V(3, N) and a rotation matrix R, the output point

* Builds the index */

Inputs: - matching_scores: Tensor of shape (B, N+1, M+1) giving the matching scores - gt_masks: Tensor of shape (B, N, M) giv

Encode node [semantic, bbox, shape] features.

background (floors, carpets) are considered. pred: (a,2), gt: (b,2) return: true_pos, false_pos

replace the f_feats in instances_knn_dict with the fused f_feats fused_feats_dict: - features: (P, C), fine-level features -

Input, - matches: tensor of shape (num_matches,2), [src_id,tar_id] - src_labels: list of string - ref_labels: list of str

convert log assignment matrix to assignment matrix

Encode all nodes in both graphs.

build assignment matrix from descriptors

r"""Sinkhorn Optimal Transport (SuperGlue style) forward. Args: scores: torch.Tensor (B, M, N) row_masks: torch.Tenso

Input, - x: (N,C)

Input: f_points: (P, 3), f_feats: (P, C), f_instances: (P, 1), instances_centroids: (N, 3)

Read a scene pair. Encode all the fine points and features.

Read a scene pair. Encode all the fine points and features.

r"""Point Matching Module forward propagation with Local-to-Global registration. There are B pairs of matched nodes. Args:

r""" encoder_hidden_states (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`): Seq

Input, - feats: (N,D) - edges: (E,2)

Input, - src_batch: (B+1) - tar_batch: (B+1)

r"""Sinusoidal Positional Embedding. Args: emb_indices: torch.Tensor (*), (N',K) or (N',K,2) Returns: embedd

r"""Sinusoidal Positional Embedding. Args: emb_indices: torch.Tensor (*), (N',K,2) Returns: embeddings: torc

Input: - x: (N, in_channels) - pos: (N, 3) - anchor_indices: (N') - corner_indices: (N',K,2)

encode position vector x: (b,n,3) return (2,b,n,d/2)

q: (B,K,D)