Open3D (C++ API)  0.18.0
Public Member Functions | Static Public Member Functions
open3d::t::geometry::VoxelBlockGrid Class Reference

#include <VoxelBlockGrid.h>

Public Member Functions

 VoxelBlockGrid ()=default
 
 VoxelBlockGrid (const std::vector< std::string > &attr_names, const std::vector< core::Dtype > &attr_dtypes, const std::vector< core::SizeVector > &attr_channels, float voxel_size=0.0058, int64_t block_resolution=16, int64_t block_count=10000, const core::Device &device=core::Device("CPU:0"), const core::HashBackendType &backend=core::HashBackendType::Default)
 Default Constructor. Example: VoxelBlockGrid({"tsdf", "weight", "color"}, {core::Float32, core::UInt16, core::UInt16}, {{1}, {1}, {3}}, 0.005, 16, 10000, core::Device("CUDA:0"), core::HashBackendType::Default);. More...
 
core::HashMap GetHashMap ()
 
core::Tensor GetAttribute (const std::string &attr_name) const
 
core::Tensor GetVoxelIndices (const core::Tensor &buf_indices) const
 
core::Tensor GetVoxelIndices () const
 Get all active voxel indices. More...
 
core::Tensor GetVoxelCoordinates (const core::Tensor &voxel_indices) const
 
std::pair< core::Tensor, core::Tensor > GetVoxelCoordinatesAndFlattenedIndices (const core::Tensor &buf_indices)
 
std::pair< core::Tensor, core::Tensor > GetVoxelCoordinatesAndFlattenedIndices ()
 
core::Tensor GetUniqueBlockCoordinates (const Image &depth, const core::Tensor &intrinsic, const core::Tensor &extrinsic, float depth_scale=1000.0f, float depth_max=3.0f, float trunc_voxel_multiplier=8.0)
 
core::Tensor GetUniqueBlockCoordinates (const PointCloud &pcd, float trunc_voxel_multiplier=8.0)
 Obtain active block coordinates from a point cloud. More...
 
void Integrate (const core::Tensor &block_coords, const Image &depth, const Image &color, const core::Tensor &depth_intrinsic, const core::Tensor &color_intrinsic, const core::Tensor &extrinsic, float depth_scale=1000.0f, float depth_max=3.0f, float trunc_voxel_multiplier=8.0f)
 
void Integrate (const core::Tensor &block_coords, const Image &depth, const Image &color, const core::Tensor &intrinsic, const core::Tensor &extrinsic, float depth_scale=1000.0f, float depth_max=3.0f, float trunc_voxel_multiplier=8.0f)
 
void Integrate (const core::Tensor &block_coords, const Image &depth, const core::Tensor &intrinsic, const core::Tensor &extrinsic, float depth_scale=1000.0f, float depth_max=3.0f, float trunc_voxel_multiplier=8.0f)
 
TensorMap RayCast (const core::Tensor &block_coords, const core::Tensor &intrinsic, const core::Tensor &extrinsic, int width, int height, const std::vector< std::string > attrs={"depth", "color"}, float depth_scale=1000.0f, float depth_min=0.1f, float depth_max=3.0f, float weight_threshold=3.0f, float trunc_voxel_multiplier=8.0f, int range_map_down_factor=8)
 
PointCloud ExtractPointCloud (float weight_threshold=3.0f, int estimated_point_number=-1)
 
TriangleMesh ExtractTriangleMesh (float weight_threshold=3.0f, int estimated_vertex_numer=-1)
 
void Save (const std::string &file_name) const
 Save a voxel block grid to a .npz file. More...
 
VoxelBlockGrid To (const core::Device &device, bool copy=false) const
 Convert the hash map to another device. More...
 

Static Public Member Functions

static VoxelBlockGrid Load (const std::string &file_name)
 Load a voxel block grid from a .npz file. More...
 

Detailed Description

A voxel block grid is a sparse grid of voxel blocks. Each voxel block is a dense 3D array, preserving local data distribution. If the block_resolution is set to 1, then the VoxelBlockGrid degenerates to a sparse voxel grid.

Constructor & Destructor Documentation

◆ VoxelBlockGrid() [1/2]

open3d::t::geometry::VoxelBlockGrid::VoxelBlockGrid ( )
default

◆ VoxelBlockGrid() [2/2]

open3d::t::geometry::VoxelBlockGrid::VoxelBlockGrid ( const std::vector< std::string > &  attr_names,
const std::vector< core::Dtype > &  attr_dtypes,
const std::vector< core::SizeVector > &  attr_channels,
float  voxel_size = 0.0058,
int64_t  block_resolution = 16,
int64_t  block_count = 10000,
const core::Device &  device = core::Device("CPU:0"),
const core::HashBackendType &  backend = core::HashBackendType::Default 
)

Default Constructor. Example: VoxelBlockGrid({"tsdf", "weight", "color"}, {core::Float32, core::UInt16, core::UInt16}, {{1}, {1}, {3}}, 0.005, 16, 10000, core::Device("CUDA:0"), core::HashBackendType::Default);.

Member Function Documentation

◆ ExtractPointCloud()

PointCloud open3d::t::geometry::VoxelBlockGrid::ExtractPointCloud ( float  weight_threshold = 3.0f,
int  estimated_point_number = -1 
)

Specific operation for TSDF volumes. Extract point cloud at isosurface points. Weight threshold is used to filter outliers. By default we use 3.0, where we assume a reliable surface point comes from the fusion of at least 3 viewpoints. Use as low as 0.0 to accept all the possible observations. Estimated point numbers optionally speeds up the process by a one-pass extraction with pre-allocated buffers. Use -1 when no estimate is available.

◆ ExtractTriangleMesh()

TriangleMesh open3d::t::geometry::VoxelBlockGrid::ExtractTriangleMesh ( float  weight_threshold = 3.0f,
int  estimated_vertex_numer = -1 
)

Specific operation for TSDF volumes. Extract mesh near iso-surfaces with Marching Cubes. Weight threshold is used to filter outliers. By default we use 3.0, where we assume a reliable surface point comes from the fusion of at least 3 viewpoints. Use as low as 0.0 to accept all the possible observations. Estimated point numbers optionally speeds up the process by a one-pass extraction with pre-allocated buffers. Use -1 when no estimate is available.

◆ GetAttribute()

core::Tensor open3d::t::geometry::VoxelBlockGrid::GetAttribute ( const std::string &  attr_name) const

Get the attribute tensor corresponding to the attribute name. A sugar for hashmap.GetValueTensor(i)

◆ GetHashMap()

core::HashMap open3d::t::geometry::VoxelBlockGrid::GetHashMap ( )
inline

Get the underlying hash map that stores values in structure of arrays (SoA).

◆ GetUniqueBlockCoordinates() [1/2]

core::Tensor open3d::t::geometry::VoxelBlockGrid::GetUniqueBlockCoordinates ( const Image &  depth,
const core::Tensor &  intrinsic,
const core::Tensor &  extrinsic,
float  depth_scale = 1000.0f,
float  depth_max = 3.0f,
float  trunc_voxel_multiplier = 8.0 
)

Get a (3, M) active block coordinates from a depth image, with potential duplicates removed. Note: these coordinates are not activated in the internal sparse voxel block. They need to be inserted in the hash map.

◆ GetUniqueBlockCoordinates() [2/2]

core::Tensor open3d::t::geometry::VoxelBlockGrid::GetUniqueBlockCoordinates ( const PointCloud &  pcd,
float  trunc_voxel_multiplier = 8.0 
)

Obtain active block coordinates from a point cloud.

◆ GetVoxelCoordinates()

core::Tensor open3d::t::geometry::VoxelBlockGrid::GetVoxelCoordinates ( const core::Tensor &  voxel_indices) const

Get a (3, hashmap.Size() * resolution^3) coordinate tensor of active voxels per block, used for geometry transformation jointly with indices from GetVoxelIndices.

Example: For a voxel block grid with (2, 2, 2) block resolution, if the active block coordinates are {(-1, 3, 2), (0, 2, 4)}, the returned result will be a (3, 2 x 8) tensor given by: { key_tensor[voxel_indices[0]] * block_resolution_ + voxel_indices[1] key_tensor[voxel_indices[0]] * block_resolution_ + voxel_indices[2] key_tensor[voxel_indices[0]] * block_resolution_ + voxel_indices[3] } Note: the coordinates are VOXEL COORDINATES in Int64. To access metric coordinates, multiply by voxel size.

◆ GetVoxelCoordinatesAndFlattenedIndices() [1/2]

std::pair< core::Tensor, core::Tensor > open3d::t::geometry::VoxelBlockGrid::GetVoxelCoordinatesAndFlattenedIndices ( )

Same as above, but N is the number of all the active voxels with blocks stored in the hash map.

◆ GetVoxelCoordinatesAndFlattenedIndices() [2/2]

std::pair< core::Tensor, core::Tensor > open3d::t::geometry::VoxelBlockGrid::GetVoxelCoordinatesAndFlattenedIndices ( const core::Tensor &  buf_indices)

Accelerated combination of GetVoxelIndices and GetVoxelCoordinates. Returns a (N, 3) coordinate in float, and a (N, ) flattened index tensor, where N is the number of active voxels located at buf_indices.

◆ GetVoxelIndices() [1/2]

core::Tensor open3d::t::geometry::VoxelBlockGrid::GetVoxelIndices ( ) const

Get all active voxel indices.

◆ GetVoxelIndices() [2/2]

core::Tensor open3d::t::geometry::VoxelBlockGrid::GetVoxelIndices ( const core::Tensor &  buf_indices) const

Get a (4, N), Int64 index tensor for active voxels, used for advanced indexing. Returned index tensor can access selected value buffers in order of (buf_index, index_voxel_x, index_voxel_y, index_voxel_z).

Example: For a voxel block grid with (2, 2, 2) block resolution, if the active block coordinates are at buffer index {(2, 4)} given by GetActiveIndices() from the underlying hash map, the returned result will be a (4, 2 x 8) tensor: { (2, 0, 0, 0), (2, 1, 0, 0), (2, 0, 1, 0), (2, 1, 1, 0), (2, 0, 0, 1), (2, 1, 0, 1), (2, 0, 1, 1), (2, 1, 1, 1), (4, 0, 0, 0), (4, 1, 0, 0), (4, 0, 1, 0), (4, 1, 1, 0), (4, 0, 0, 1), (4, 1, 0, 1), (4, 0, 1, 1), (4, 1, 1, 1), } Note: the slicing order is z-y-x.

TODO: Check if kernel version is necessary.

operator % is not supported now

◆ Integrate() [1/3]

void open3d::t::geometry::VoxelBlockGrid::Integrate ( const core::Tensor &  block_coords,
const Image &  depth,
const core::Tensor &  intrinsic,
const core::Tensor &  extrinsic,
float  depth_scale = 1000.0f,
float  depth_max = 3.0f,
float  trunc_voxel_multiplier = 8.0f 
)

Specific operation for TSDF volumes. Similar to RGB-D integration, but only applied to depth.

◆ Integrate() [2/3]

void open3d::t::geometry::VoxelBlockGrid::Integrate ( const core::Tensor &  block_coords,
const Image &  depth,
const Image &  color,
const core::Tensor &  depth_intrinsic,
const core::Tensor &  color_intrinsic,
const core::Tensor &  extrinsic,
float  depth_scale = 1000.0f,
float  depth_max = 3.0f,
float  trunc_voxel_multiplier = 8.0f 
)

Specific operation for TSDF volumes. Integrate an RGB-D frame in the selected block coordinates using pinhole camera model. For built-in kernels, we support efficient hash map types for SLAM: tsdf: float, weight: uint16_t, color: uint16_t and accurate mode for differentiable rendering: tsdf/weight/color: float We assume input data are either raw: depth: uint16_t, color: uint8_t or depth/color: float. To support other types and properties, users should combine GetUniqueBlockCoordinates, GetVoxelIndices, and GetVoxelCoordinates, with self-defined operations.

◆ Integrate() [3/3]

void open3d::t::geometry::VoxelBlockGrid::Integrate ( const core::Tensor &  block_coords,
const Image &  depth,
const Image &  color,
const core::Tensor &  intrinsic,
const core::Tensor &  extrinsic,
float  depth_scale = 1000.0f,
float  depth_max = 3.0f,
float  trunc_voxel_multiplier = 8.0f 
)

Specific operation for TSDF volumes. Similar to RGB-D integration, but uses the same intrinsics for depth and color.

◆ Load()

VoxelBlockGrid open3d::t::geometry::VoxelBlockGrid::Load ( const std::string &  file_name)
static

Load a voxel block grid from a .npz file.

◆ RayCast()

TensorMap open3d::t::geometry::VoxelBlockGrid::RayCast ( const core::Tensor &  block_coords,
const core::Tensor &  intrinsic,
const core::Tensor &  extrinsic,
int  width,
int  height,
const std::vector< std::string >  attrs = {"depth", "color"},
float  depth_scale = 1000.0f,
float  depth_min = 0.1f,
float  depth_max = 3.0f,
float  weight_threshold = 3.0f,
float  trunc_voxel_multiplier = 8.0f,
int  range_map_down_factor = 8 
)

Specific operation for TSDF volumes. Perform volumetric ray casting in the selected block coordinates. Return selected properties from the frame. Supported attributes: Conventional rendering: vertex, depth, color, normal, range Differentiable rendering (voxel-wise): mask, index, (interpolation) ratio. The block coordinates in the frustum can be taken from GetUniqueBlockCoordinates. All the block coordinates can be taken from GetHashMap().GetKeyTensor().

◆ Save()

void open3d::t::geometry::VoxelBlockGrid::Save ( const std::string &  file_name) const

Save a voxel block grid to a .npz file.

◆ To()

VoxelBlockGrid open3d::t::geometry::VoxelBlockGrid::To ( const core::Device &  device,
bool  copy = false 
) const

Convert the hash map to another device.

The documentation for this class was generated from the following files:
Generated by doxygen 1.9.1