Converts a pd.DataFrame to a [N, 4, 4] array of poses read from a dataframe Parameters ---------- df : pd.DataFrame A DataFr

Computes the new pose and stores it in memory Args: data_dict (dict): The new frame's data as a dictionary returned by t

Save the odometry, loop and absolute constraints on disk

Converts a ECEF GPS position to a LLU frame

Computes an Estimate of timestamps for rotating lasers Each point are expressed in spherical coordinates, The timestamps are assigned based o

Launches the evaluation for an epoch

Applies a filter which modifies the state of the data_dict

Finds the index of the first point with azimuth equal to azimuth_bin After at least min_num_points have been passed

Estimates the ground truth poses for the LiDAR frames The output files are saved as text file under self.root_dir / <sequence_name>.poses

Returns the ground truth for the dataset_config for a given sequence

Returns the computed relative poses along the trajectory

Returns: (train_dataset, eval_dataset, test_dataset) : A tuple of `DatasetOfSequences`

Returns the vertex index in the pose graph of the gps vertex corresponding to a pose index

Sample one point per voxel hash, returns the sample points and the indices of the sampled points Args: pointcloud (np.ndarray): The input

Clears and Initialize the Local Map

Loads a checkpoint file saved during training The checkpoint file is a python dictionary saved, The dictionary contains the

Returns the .train, validation and test datasets as options

Returns the loss module for the specific trainer Returns ------- nn.Module The loss module computes the

Main function: builds iteratively the UrbanLoco ground truth files

Returns the maximum number of workers allowed by this dataset Note: Not respecting this constraint can lead to undefined behaviour for

Set all pixel data of a projection map which have a nan to a default value

Finds nearest neighbors correspondences in the map for a set of points

Returns a neighborhood convolution kernel, It is a tensor which can be used to extract the K * K points in the neighborhood of a give

Processes a next frame by adding all trajectory constraints to the graph slam

Initializes the Algorithm ()

Extract velodyne packet_ids from the array of ring ids (which can be used to simulate timestamps)

Parses arguments from the command line

Computes the Point-to-Plane loss between a target vertex map and a reference Parameters ---------- vm_target: torch.

Args: data_dict (dict): The new frame (consisting of a dictionary of data items) returned by the Dataset

Computes the new pose and stores it in memory Args: data_dict (dict): The new frame (consisting of a dictionary of data

Default SphericalProjetor for KITTI (projection of a pointcloud into a Vertex Map)

Default SphericalProjetor for NHCD (projection of a pointcloud into a Vertex Map)

Read a timestamps file and convert it to float64 values

Registers a new frame against the Local Map Args: target_points (torch.Tensor): The target Ver initial_estim

Returns the registered absolute constraints

Returns the registered loop constraints

Returns the registered odometry constraints

Returns the relative poses between two consecutive frames corrected after global optimization

The main entry point to the script running the SLAM

Rescales the values of an image, for more convenient visualization Gray Image will have values in [0, 1] RGB Images will have values in [

Returns the point to plane residuals

Returns the point to point residuals

Builds a pose matrix tensor from its rotation parameters Parameters ---------- rot_params : [B, ROT_P] T

Returns ------- torch.Tensor A [B, P] tensor with the parameters of the representation of the rotation matrices

Runs the LiDAR Odometry algorithm on the different datasets

The main entry point to the script running the SLAM

Sample one point per hash value, returns the sample points and the indices of the sampled points Args: pointcloud (np.ndarray): The input

Returns the points sampled

Saves the modules and optimizer parameters in a checkpoint file

Saves the config to Disk

Saves the real new motion into the algorithm

Returns a key which defines a relative constraint between two poses

Returns a set of constraints read from `data_dict`

Sends data to the device if it can torch.Tensor are sent to the device, containers send all the torch.Tensor to the devices Other da

Reads projection params from the arguments and set the appropriate parameters All named arguments are optional, and will only be set

Returns True if the sizes matches the tensor shape

Returns ------- [B, 3, 3, 3]

Parameters ---------- pose_params : The pose parameters with euler convention [B, 6] [B, :3] the translation parameters

Launches the training for an epoch

Updates trajectory Args: trajectory (np.ndarray): The absolute poses making the trajectory `(N, 4, 4)`

Visualize the images in data_dict in OpenCV windows The data_dict must have all keys in in self.channels Each it

Computes a hash given the int coordinates of a 3D voxel Args: x (np.int64): the x coordinate of a voxel y (np.int64): the y

Returns the weighted tensor of residuals given a tensor of initial unweighted residuals Args: residuals (torch.Tensor):