Port pointcloud creation to numpy. (#175)

* Port pointcloud creation to numpy.

Signed-off-by: Florian Vahl <[email protected]>

* Add support for differently typed fields.

Signed-off-by: Florian Vahl <[email protected]>

* Add support for reading structured arrays

Signed-off-by: Florian Vahl <[email protected]>

* Adapt point cloud creation to utilize organized point clouds and multiple field datatypes

Signed-off-by: Florian Vahl <[email protected]>

* Format imports

Signed-off-by: Florian Vahl <[email protected]>

* Apply autopep8

Signed-off-by: Florian Vahl <[email protected]>

* Add python tmp files to .gitignore

Signed-off-by: Florian Vahl <[email protected]>

* Fix docsting test

Signed-off-by: Florian Vahl <[email protected]>

* Apply flake8 style

Signed-off-by: Florian Vahl <[email protected]>

* Add list cast to points. \n They are tuples otherwise, which lets the check fail because list!=tuple

Signed-off-by: Florian Vahl <[email protected]>

* Add more tests

Signed-off-by: Florian Vahl <[email protected]>

* Fix issues discorvered while testing

Signed-off-by: Florian Vahl <[email protected]>

* Fix typo

Signed-off-by: Florian Vahl <[email protected]>

* Fix docstring

Signed-off-by: Florian Vahl <[email protected]>

* Fix docstring

Signed-off-by: Florian Vahl <[email protected]>

* Fix grammar

Signed-off-by: Florian Vahl <[email protected]>

* Add support for fields with count > 1

Signed-off-by: Florian Vahl <[email protected]>

* Use casting suggested by apockill

Signed-off-by: Florian Vahl <[email protected]>

* Fix flake8 errors

Signed-off-by: Florian Vahl <[email protected]>

* Fix import order

Signed-off-by: Florian Vahl <[email protected]>

Author: Flova

.gitignore

@@ -26,3 +26,7 @@
 *.exe
 *.out
 *.app
+
+# Python temporary files
+*.pyc
+__pycache__/

sensor_msgs_py/package.xml

@@ -15,6 +15,7 @@
   <author email="[email protected]">Sebastian Grans</author>
 
   <exec_depend>sensor_msgs</exec_depend>
+  <exec_depend>python3-numpy</exec_depend>
 
   <test_depend>ament_copyright</test_depend>
   <test_depend>ament_flake8</test_depend>

sensor_msgs_py/sensor_msgs_py/point_cloud2.py

@@ -37,27 +37,37 @@
 sensor_msgs/src/sensor_msgs/point_cloud2.py
 """
 
+import array
 from collections import namedtuple
-import ctypes
-import math
-import struct
 import sys
+from typing import Iterable, List, NamedTuple, Optional
 
+import numpy as np
+from numpy.lib.recfunctions import (structured_to_unstructured,
+                                    unstructured_to_structured)
 from sensor_msgs.msg import PointCloud2, PointField
+from std_msgs.msg import Header
 
 
 _DATATYPES = {}
-_DATATYPES[PointField.INT8] = ('b', 1)
-_DATATYPES[PointField.UINT8] = ('B', 1)
-_DATATYPES[PointField.INT16] = ('h', 2)
-_DATATYPES[PointField.UINT16] = ('H', 2)
-_DATATYPES[PointField.INT32] = ('i', 4)
-_DATATYPES[PointField.UINT32] = ('I', 4)
-_DATATYPES[PointField.FLOAT32] = ('f', 4)
-_DATATYPES[PointField.FLOAT64] = ('d', 8)
-
-
-def read_points(cloud, field_names=None, skip_nans=False, uvs=[]):
+_DATATYPES[PointField.INT8] = np.dtype(np.int8)
+_DATATYPES[PointField.UINT8] = np.dtype(np.uint8)
+_DATATYPES[PointField.INT16] = np.dtype(np.int16)
+_DATATYPES[PointField.UINT16] = np.dtype(np.uint16)
+_DATATYPES[PointField.INT32] = np.dtype(np.int32)
+_DATATYPES[PointField.UINT32] = np.dtype(np.uint32)
+_DATATYPES[PointField.FLOAT32] = np.dtype(np.float32)
+_DATATYPES[PointField.FLOAT64] = np.dtype(np.float64)
+
+DUMMY_FIELD_PREFIX = 'unnamed_field'
+
+
+def read_points(
+        cloud: PointCloud2,
+        field_names: Optional[List[str]] = None,
+        skip_nans: bool = False,
+        uvs: Optional[Iterable] = None,
+        reshape_organized_cloud: bool = False) -> np.ndarray:
     """
     Read points from a sensor_msgs.PointCloud2 message.
 
@@ -67,56 +77,92 @@ def read_points(cloud, field_names=None, skip_nans=False, uvs=[]):
     :param skip_nans: If True, then don't return any point with a NaN value.
                       (Type: Bool, Default: False)
     :param uvs: If specified, then only return the points at the given
-        coordinates. (Type: Iterable, Default: empty list)
-    :return: Generator which yields a list of values for each point.
+        coordinates. (Type: Iterable, Default: None)
+    :param reshape_organized_cloud: Returns the array as an 2D organized point cloud if set.
+    :return: Structured NumPy array containing all points.
     """
     assert isinstance(cloud, PointCloud2), \
-        'cloud is not a sensor_msgs.msg.PointCloud2'
-    fmt = _get_struct_fmt(cloud.is_bigendian, cloud.fields, field_names)
-    width, height, point_step, row_step, data, isnan = \
-        cloud.width, cloud.height, \
-        cloud.point_step, cloud.row_step, \
-        cloud.data, math.isnan
-
-    unpack_from = struct.Struct(fmt).unpack_from
-
-    if skip_nans:
-        if uvs:
-            for u, v in uvs:
-                p = unpack_from(data, (row_step * v) + (point_step * u))
-                has_nan = False
-                for pv in p:
-                    if isnan(pv):
-                        has_nan = True
-                        break
-                if not has_nan:
-                    yield p
-        else:
-            for v in range(height):
-                offset = row_step * v
-                for u in range(width):
-                    p = unpack_from(data, offset)
-                    has_nan = False
-                    for pv in p:
-                        if isnan(pv):
-                            has_nan = True
-                            break
-                    if not has_nan:
-                        yield p
-                    offset += point_step
-    else:
-        if uvs:
-            for u, v in uvs:
-                yield unpack_from(data, (row_step * v) + (point_step * u))
-        else:
-            for v in range(height):
-                offset = row_step * v
-                for u in range(width):
-                    yield unpack_from(data, offset)
-                    offset += point_step
+        'Cloud is not a sensor_msgs.msg.PointCloud2'
+
+    # Cast bytes to numpy array
+    points = np.ndarray(
+        shape=(cloud.width * cloud.height, ),
+        dtype=dtype_from_fields(cloud.fields),
+        buffer=cloud.data)
+
+    # Keep only the requested fields
+    if field_names is not None:
+        assert all(field_name in points.dtype.names for field_name in field_names), \
+            'Requests field is not in the fields of the PointCloud!'
+        # Mask fields
+        points = points[list(field_names)]
+
+    # Swap array if byte order does not match
+    if bool(sys.byteorder != 'little') != bool(cloud.is_bigendian):
+        points = points.byteswap(inplace=True)
+
+    # Check if we want to drop points with nan values
+    if skip_nans and not cloud.is_dense:
+        # Init mask which selects all points
+        not_nan_mask = np.ones(len(points), dtype=bool)
+        for field_name in points.dtype.names:
+            # Only keep points without any non values in the mask
+            not_nan_mask = np.logical_and(
+                not_nan_mask, ~np.isnan(points[field_name]))
+        # Select these points
+        points = points[not_nan_mask]
+
+    # Select points indexed by the uvs field
+    if uvs is not None:
+        # Don't convert to numpy array if it is already one
+        if not isinstance(uvs, np.ndarray):
+            uvs = np.fromiter(uvs, int)
+        # Index requested points
+        points = points[uvs]
+
+    # Cast into 2d array if cloud is 'organized'
+    if reshape_organized_cloud and cloud.height > 1:
+        points = points.reshape(cloud.width, cloud.height)
+
+    return points
+
+
+def read_points_numpy(
+        cloud: PointCloud2,
+        field_names: Optional[List[str]] = None,
+        skip_nans: bool = False,
+        uvs: Optional[Iterable] = None,
+        reshape_organized_cloud: bool = False) -> np.ndarray:
+    """
+    Read equally typed fields from sensor_msgs.PointCloud2 message as a unstructured numpy array.
 
+    This method is better suited if one wants to perform math operations
+    on e.g. all x,y,z fields.
+    But it is limited to fields with the same dtype as unstructured numpy arrays
+    only contain one dtype.
 
-def read_points_list(cloud, field_names=None, skip_nans=False, uvs=[]):
+    :param cloud: The point cloud to read from sensor_msgs.PointCloud2.
+    :param field_names: The names of fields to read. If None, read all fields.
+                        (Type: Iterable, Default: None)
+    :param skip_nans: If True, then don't return any point with a NaN value.
+                      (Type: Bool, Default: False)
+    :param uvs: If specified, then only return the points at the given
+        coordinates. (Type: Iterable, Default: None)
+    :param reshape_organized_cloud: Returns the array as an 2D organized point cloud if set.
+    :return: Numpy array containing all points.
+    """
+    assert all(cloud.fields[0].datatype == field.datatype for field in cloud.fields[1:]), \
+        'All fields need to have the same datatype. Use `read_points()` otherwise.'
+    structured_numpy_array = read_points(
+        cloud, field_names, skip_nans, uvs, reshape_organized_cloud)
+    return structured_to_unstructured(structured_numpy_array)
+
+
+def read_points_list(
+        cloud: PointCloud2,
+        field_names: Optional[List[str]] = None,
+        skip_nans: bool = False,
+        uvs: Optional[Iterable] = None) -> List[NamedTuple]:
     """
     Read points from a sensor_msgs.PointCloud2 message.
 
@@ -129,7 +175,7 @@ def read_points_list(cloud, field_names=None, skip_nans=False, uvs=[]):
     :param skip_nans: If True, then don't return any point with a NaN value.
                       (Type: Bool, Default: False)
     :param uvs: If specified, then only return the points at the given
-                coordinates. (Type: Iterable, Default: empty list]
+                coordinates. (Type: Iterable, Default: None]
     :return: List of namedtuples containing the values for each point
     """
     assert isinstance(cloud, PointCloud2), \
@@ -144,7 +190,53 @@ def read_points_list(cloud, field_names=None, skip_nans=False, uvs=[]):
                                                 skip_nans, uvs)]
 
 
-def create_cloud(header, fields, points):
+def dtype_from_fields(fields: Iterable[PointField]) -> np.dtype:
+    """
+    Convert a Iterable of sensor_msgs.msg.PointField messages to a np.dtype.
+
+    :param fields: The point cloud fields.
+                   (Type: iterable of sensor_msgs.msg.PointField)
+    :returns: NumPy datatype
+    """
+    # Create a lists containing the names, offsets and datatypes of all fields
+    field_names = []
+    field_offsets = []
+    field_datatypes = []
+    for i, field in enumerate(fields):
+        # Datatype as numpy datatype
+        datatype = _DATATYPES[field.datatype]
+        # Name field
+        if field.name == '':
+            name = f'{DUMMY_FIELD_PREFIX}_{i}'
+        else:
+            name = field.name
+        # Handle fields with count > 1 by creating subfields with a suffix consiting
+        # of "_" followed by the subfield counter [0 -> (count - 1)]
+        assert field.count > 0, "Can't process fields with count = 0."
+        for a in range(field.count):
+            # Add suffix if we have multiple subfields
+            if field.count > 1:
+                subfield_name = f'{name}_{a}'
+            else:
+                subfield_name = name
+            assert subfield_name not in field_names, 'Duplicate field names are not allowed!'
+            field_names.append(subfield_name)
+            # Create new offset that includes subfields
+            field_offsets.append(field.offset + a * datatype.itemsize)
+            field_datatypes.append(datatype.str)
+
+    # Create a tuple for each field containing name and data type
+    return np.dtype({
+        'names': field_names,
+        'formats': field_datatypes,
+        'offsets': field_offsets,
+    })
+
+
+def create_cloud(
+        header: Header,
+        fields: Iterable[PointField],
+        points: Iterable) -> PointCloud2:
     """
     Create a sensor_msgs.msg.PointCloud2 message.
 
@@ -157,28 +249,61 @@ def create_cloud(header, fields, points):
                    the fields parameter)
     :return: The point cloud as sensor_msgs.msg.PointCloud2
     """
-    cloud_struct = struct.Struct(_get_struct_fmt(False, fields))
-
-    buff = ctypes.create_string_buffer(cloud_struct.size * len(points))
-
-    point_step, pack_into = cloud_struct.size, cloud_struct.pack_into
-    offset = 0
-    for p in points:
-        pack_into(buff, offset, *p)
-        offset += point_step
-
-    return PointCloud2(header=header,
-                       height=1,
-                       width=len(points),
-                       is_dense=False,
-                       is_bigendian=False,
-                       fields=fields,
-                       point_step=cloud_struct.size,
-                       row_step=cloud_struct.size * len(points),
-                       data=buff.raw)
-
-
-def create_cloud_xyz32(header, points):
+    # Check if input is numpy array
+    if isinstance(points, np.ndarray):
+        # Check if this is an unstructured array
+        if points.dtype.names is None:
+            assert all(fields[0].datatype == field.datatype for field in fields[1:]), \
+                'All fields need to have the same datatype. Pass a structured NumPy array \
+                    with multiple dtypes otherwise.'
+            # Convert unstructured to structured array
+            points = unstructured_to_structured(
+                points,
+                dtype=dtype_from_fields(fields))
+        else:
+            assert points.dtype == dtype_from_fields(fields), \
+                'PointFields and structured NumPy array dtype do not match for all fields! \
+                    Check their field order, names and types.'
+    else:
+        # Cast python objects to structured NumPy array (slow)
+        points = np.array(
+            # Points need to be tuples in the structured array
+            list(map(tuple, points)),
+            dtype=dtype_from_fields(fields))
+
+    # Handle organized clouds
+    assert len(points.shape) <= 2, \
+        'Too many dimensions for organized cloud! \
+            Points can only be organized in max. two dimensional space'
+    height = 1
+    width = points.shape[0]
+    # Check if input points are an organized cloud (2D array of points)
+    if len(points.shape) == 2:
+        height = points.shape[1]
+
+    # Convert numpy points to array.array
+    memory_view = memoryview(points)
+    casted = memory_view.cast('B')
+    array_array = array.array('B')
+    array_array.frombytes(casted)
+
+    # Put everything together
+    cloud = PointCloud2(
+        header=header,
+        height=height,
+        width=width,
+        is_dense=False,
+        is_bigendian=sys.byteorder != 'little',
+        fields=fields,
+        point_step=points.dtype.itemsize,
+        row_step=(points.dtype.itemsize * width))
+    # Set cloud via property instead of the constructor because of the bug described in
+    # https://github.com/ros2/common_interfaces/issues/176
+    cloud.data = array_array
+    return cloud
+
+
+def create_cloud_xyz32(header: Header, points: Iterable) -> PointCloud2:
     """
     Create a sensor_msgs.msg.PointCloud2 message with (x, y, z) fields.
 
@@ -193,23 +318,3 @@ def create_cloud_xyz32(header, points):
               PointField(name='z', offset=8,
                          datatype=PointField.FLOAT32, count=1)]
     return create_cloud(header, fields, points)
-
-
-def _get_struct_fmt(is_bigendian, fields, field_names=None):
-    fmt = '>' if is_bigendian else '<'
-
-    offset = 0
-    for field in (f for f in sorted(fields, key=lambda f: f.offset)
-                  if field_names is None or f.name in field_names):
-        if offset < field.offset:
-            fmt += 'x' * (field.offset - offset)
-            offset = field.offset
-        if field.datatype not in _DATATYPES:
-            print('Skipping unknown PointField datatype [%d]' %
-                  field.datatype, file=sys.stderr)
-        else:
-            datatype_fmt, datatype_length = _DATATYPES[field.datatype]
-            fmt += field.count * datatype_fmt
-            offset += field.count * datatype_length
-
-    return fmt