Inherit from torch_geometric.Data with extensions tailored to our specific needs.

Holder for a Nested Acyclic Graph, containing a list of nested partitions of the same point cloud.

Child class of CSRData to simplify some common operations dedicated to cluster-point indexing.

Child class of CSRData to simplify some common operations dedicated to instance labels clustering. In particular, this data structure stores t

Adds the index of the nodes to the Data object attributes. This allows tracking nodes from the output back to the input Data object.

A simple holder for panoptic segmentation model output, with a few helper methods for manipulating the predictions and targets (if any).

Clusters 3D points into voxels with size :attr:`size`. By default, some special keys undergo dedicated grouping mechanisms. The `_VOTING_KEY

Remove attributes of a NAG object based on their name. :param level: int or str Level at which to remove attributes. Can be an int or a s

Get attributes from their keys and concatenate them to x. :param keys: str or list(str) The feature concatenated to 'to' :param to: s

Cast Data attributes to the provided integer and floating point dtypes. In case 'rgb' or 'mean_rgb' is found, `rgb_to_float` will decide wheth

Apply some random contrast to the point colors. credit: https://github.com/guochengqian/openpoints :param p: float Probability of th

Randomly set point colors to 0. :param p: float Probability of the transform to be applied :param x_idx: int If specified, th

Normalize the colors using given means and standard deviations. credit: https://github.com/guochengqian/openpoints :param mean: list

r"""Creates a criterion that measures Learned Perceptual Image Patch Similarity (LPIPS). Arguments: net_type (str): the network type

Compute the non-oriented graph used for instance and panoptic segmentation. We choose the following assignment rule: - each superpoint

Remove attributes of a Data object based on their name. :param keys: str of list(str) List of attribute names :param strict: bool

Select columns of an attribute based on their indices. :param key: str The Data attribute whose columns should be selected :param idx

Shuffle the order of points in a Data object.

Create the adjacency graph in `edge_index` and `edge_attr` based on the `Data.neighbor_index` and `Data.neighbor_distance`. NB: the produced

Base class for sampling subgraphs from a NAG. It randomly picks `k` seed nodes from `i_level`, from which `k` subgraphs can be grown. Child cl

Inherit from torch_geometric.Batch with extensions tailored to our specific needs. NB: contrary to PyGBatch's dynamic inheritance behavior, w

Wrapper class of CSRData to build a batch from a list of CSRData data and reconstruct it afterward. When defining a subclass A of CSRDat

Implements the CSRData format and associated mechanisms in Torch. When defining a subclass A of CSRData, it is recommended to create an assoc

Center the position of all nodes of all levels of a NAG around their level-0 centroid.

Wrapper for Cluster batching.

Parent class for color-based point Transforms, to avoid redundant code. :param x_idx: int If specified, the colors will be searched i

Creates edges for isolated nodes. Each isolated node is connected to the `k` nearest nodes. If the Data graph contains edge features in `Data.

Partition a graph contained in a `Data` object using cut-pursuit. The input `Data` object is assumed to hold the following attributes: - `p

Move Data object to specified device.

Convert Data to a single-level NAG.

Compute horizontal edges for all NAG levels except its first (i.e. the 0-level). These are the edges connecting the segments at each level, eq

Randomly set a Data attribute column to 0. :param p: float Probability of a column to be dropped :param key: str The Data att

Randomly set a Data attribute rows to 0. :param p: float Probability of a row to be dropped :param key: str The Data attribut

XY-grid-based hierarchical partition of Data. The nodes are aggregated based on their coordinates in a grid of step `size`. :param size: int

Compute pointwise elevation with respect to the ground. We do so in a two-step process where we first remove as many potentially non-ground p

Search for points with `k_min` OR MORE neighbors within a radius of `r_max`. Since removing outliers may cause some points to become outliers

Wrapper for InstanceData batching. Importantly, although instance labels in 'obj' will be updated to avoid collisions between the different ba

K-NN search for each point in Data. Neighbors and corresponding distances are stored in `Data.neighbor_index` and `Data.neighbor_distance`, r

A simple class to be used for learnable parameters (e.g. learnable position encodings, queries, keys, ...). Using this is useful to use custom

Get attributes from their keys and concatenate them to x. :param level: int or str Level at which to remove attributes. Can be an int or

Add self-loops to all NAG levels having a horizontal graph. If the edges have attributes, the self-loops will receive 0-features.

Wrapper for NAG batching.

Cast NAG attributes to the provided integer and floating point dtypes. In case 'rgb' or 'mean_rgb' is found and is not a floating point tensor

Apply some random contrast to the point colors. credit: https://github.com/guochengqian/openpoints :param level: int or str Level at

Randomly set point colors to 0. :param level: int or str Level at which to remove attributes. Can be an int or a str. If the latt

Normalize the colors using given means and standard deviations. credit: https://github.com/guochengqian/openpoints :param level: int or str

Randomly set a Data attribute column to 0. :param level: int or str Level at which to drop columns. Can be an int or a str. If th

Randomly set a Data attribute rows to 0. :param level: int or str Level at which to drop rows. Can be an int or a str. If the lat

Add some gaussian noise to Data['key'] for all data in a NAG. :param key: str The attribute on which to apply jittering :param sigma:

Compute the instances contained in each superpoint of each level, provided that the first level has an 'obj' attribute holding an `InstanceDat

Randomly sample nodes and edges to restrict their number within given limits. This is useful for stabilizing memory use of the model. :pa

SaveNodeIndex, applied to each NAG level.

Select the i-level nodes based on a key. The corresponding key is expected to exist in the i-level attributes and should hold a 1D boolean mas

Select columns of an attribute based on their indices. :param level: int or str Level at which to select attributes. Can be an int or a s

Move Data object to specified device.

Convert a single-level NAG to Data.

Compute the number of `low`-level elements are contained in each segment, at each above-level. Results are save in the `node_size` attribute o

Compute edge features "on-the-fly" for all i->j and j->i horizontal edges of the NAG levels except its first (i.e. the 0-level). Expects

Compute edge features "on-the-fly" for all vertical edges of the NAG levels. Optionally build some edge features that can be recovered from t

Search for points with LESS THAN `k_min` neighbors within a radius of `r_max`. Since removing outliers may cause some points to become outlie

A class to hold the output results of multiple partitions, when searching for the optimal partition parameter settings. Since metrics are only

Compute pointwise features based on what is already available in the Data object. All local geometric features assume the input ``Data`` has

Compute horizontal edges for all NAG levels except its first (i.e. the 0-level). These are the edges connecting the segments at each level, eq

Scales node positions by a randomly sampled factor ``s1, s2, s3`` within a given interval, *e.g.*, resulting in the following transformation m

Flip the node positions wrt one of the XYZ axes, with a specified probability. This transform is not very modular because it is intended to be