diff --git a/doc/source/whatsnew/v0.23.0.txt b/doc/source/whatsnew/v0.23.0.txt
index 3f7c4b3b0ccb7..05a25b122d1f4 100644
--- a/doc/source/whatsnew/v0.23.0.txt
+++ b/doc/source/whatsnew/v0.23.0.txt
@@ -1092,6 +1092,7 @@ Numeric
 - Bug in :class:`Series` constructor with an int or float list where specifying ``dtype=str``, ``dtype='str'`` or ``dtype='U'`` failed to convert the data elements to strings (:issue:`16605`)
 - Bug in :class:`Index` multiplication and division methods where operating with a ``Series`` would return an ``Index`` object instead of a ``Series`` object (:issue:`19042`)
 - Bug in the :class:`DataFrame` constructor in which data containing very large positive or very large negative numbers was causing ``OverflowError`` (:issue:`18584`)
+- Bug in the :meth:`DataFrame.from_records` constructor losing the dtypes of a empty NumPy record array (:issue:`20805`)
 - Bug in :class:`Index` constructor with ``dtype='uint64'`` where int-like floats were not coerced to :class:`UInt64Index` (:issue:`18400`)
 - Bug in :class:`DataFrame` flex arithmetic (e.g. ``df.add(other, fill_value=foo)``) with a ``fill_value`` other than ``None`` failed to raise ``NotImplementedError`` in corner cases where either the frame or ``other`` has length zero (:issue:`19522`)
 - Multiplication and division of numeric-dtyped :class:`Index` objects with timedelta-like scalars returns ``TimedeltaIndex`` instead of raising ``TypeError`` (:issue:`19333`)
diff --git a/pandas/core/frame.py b/pandas/core/frame.py
index b67ed9cfd2241..1696e3a910924 100644
--- a/pandas/core/frame.py
+++ b/pandas/core/frame.py
@@ -7387,7 +7387,10 @@ def _to_arrays(data, columns, coerce_float=False, dtype=None):
         if isinstance(data, np.ndarray):
             columns = data.dtype.names
             if columns is not None:
-                return [[]] * len(columns), columns
+                arrays = [np.array([], dtype=field_type)
+                          for field_type, _ in data.dtype.fields.values()]
+                return arrays, columns
+
         return [], []  # columns if columns is not None else []
     if isinstance(data[0], (list, tuple)):
         return _list_to_arrays(data, columns, coerce_float=coerce_float,
diff --git a/pandas/tests/frame/test_constructors.py b/pandas/tests/frame/test_constructors.py
index 47b7d60e3b6e8..7344b99db758c 100644
--- a/pandas/tests/frame/test_constructors.py
+++ b/pandas/tests/frame/test_constructors.py
@@ -1933,9 +1933,18 @@ def test_from_records_empty_with_nonempty_fields_gh3682(self):
 
         b = np.array([], dtype=[('id', np.int64), ('value', np.int64)])
         df = DataFrame.from_records(b, index='id')
-        tm.assert_index_equal(df.index, Index([], name='id'))
+        tm.assert_index_equal(df.index, Index([], name='id', dtype='int'))
         assert df.index.name == 'id'
 
+    def test_from_records_empty_dtypes(self):
+        # https://github.com/pandas-dev/pandas/issues/20805
+        a = np.array([(1, 2)], dtype=[('id', 'u8'), ('value', 'i8')])
+        result = DataFrame.from_records(a[:0])
+        expected = pd.DataFrame({"id": np.array([], dtype='u8'),
+                                 "value": np.array([], dtype='i8')},
+                                columns=['id', 'value'])
+        tm.assert_frame_equal(result, expected)
+
     def test_from_records_with_datetimes(self):
 
         # this may fail on certain platforms because of a numpy issue