diff --git a/mapillary_tools/geotag/camm_builder.py b/mapillary_tools/geotag/camm_builder.py
index f9729d77..521bea51 100644
--- a/mapillary_tools/geotag/camm_builder.py
+++ b/mapillary_tools/geotag/camm_builder.py
@@ -1,4 +1,4 @@
-import dataclasses
+import io
 import typing as T
 
 from .. import geo
@@ -228,17 +228,6 @@ def create_camm_trak(
     }
 
 
-@dataclasses.dataclass
-class CAMMPointReader(builder.Reader):
-    __slots__ = ("point",)
-
-    def __init__(self, point: geo.Point):
-        self.point = point
-
-    def read(self):
-        return build_camm_sample(self.point)
-
-
 def extract_points(fp: T.BinaryIO) -> T.Tuple[str, T.List[geo.Point]]:
     start_offset = fp.tell()
     points = camm_parser.extract_points(fp)
@@ -262,8 +251,8 @@ def camm_sample_generator2(points: T.Sequence[geo.Point]):
     def _f(
         fp: T.BinaryIO,
         moov_children: T.List[BoxDict],
-    ) -> T.Generator[builder.Reader, None, None]:
-        movie_timescale = builder._find_movie_timescale(moov_children)
+    ) -> T.Generator[io.IOBase, None, None]:
+        movie_timescale = builder.find_movie_timescale(moov_children)
         # make sure the precision of timedeltas not lower than 0.001 (1ms)
         media_timescale = max(1000, movie_timescale)
         camm_samples = list(convert_points_to_raw_samples(points, media_timescale))
@@ -279,7 +268,7 @@ def _f(
         moov_children.append(camm_trak)
 
         # if yield, the moov_children will not be modified
-        return (CAMMPointReader(point) for point in points)
+        return (io.BytesIO(build_camm_sample(point)) for point in points)
 
     return _f
 
@@ -287,7 +276,7 @@ def _f(
 def camm_sample_generator(
     fp: T.BinaryIO,
     moov_children: T.List[BoxDict],
-) -> T.Iterator[builder.Reader]:
+) -> T.Iterator[io.IOBase]:
     fp.seek(0)
     _, points = extract_points(fp)
     if not points:
diff --git a/mapillary_tools/geotag/io_utils.py b/mapillary_tools/geotag/io_utils.py
new file mode 100644
index 00000000..23edc50d
--- /dev/null
+++ b/mapillary_tools/geotag/io_utils.py
@@ -0,0 +1,178 @@
+import io
+import typing as T
+
+
+class ChainedIO(io.IOBase):
+    # is the chained stream seekable?
+    _streams: T.Sequence[io.IOBase]
+    # the beginning offset of the current stream
+    _begin_offset: int
+    # offset after SEEK_END
+    _offset_after_seek_end: int
+    # index of the current stream
+    _idx: int
+
+    def __init__(self, streams: T.Sequence[io.IOBase]):
+        for s in streams:
+            assert s.readable(), f"stream {s} must be readable"
+            assert s.seekable(), f"stream {s} must be seekable"
+            # required, otherwise inconsistent results when seeking back and forth
+            s.seek(0, io.SEEK_SET)
+        self._streams = streams
+        self._begin_offset = 0
+        self._offset_after_seek_end = 0
+        self._idx = 0
+
+    def _seek_next_stream(self) -> None:
+        """
+        seek to the end of the current stream, and seek to the beginning of the next stream
+        """
+        if self._idx < len(self._streams):
+            s = self._streams[self._idx]
+            ssize = s.seek(0, io.SEEK_END)
+
+            # update index
+            self._idx += 1
+
+            # seek to the beginning of the next stream
+            if self._idx < len(self._streams):
+                self._streams[self._idx].seek(0, io.SEEK_SET)
+
+            # update offset
+            self._begin_offset += ssize
+
+    def read(self, n: int = -1) -> bytes:
+        acc = []
+
+        while self._idx < len(self._streams):
+            data = self._streams[self._idx].read(n)
+            acc.append(data)
+            if n == -1:
+                self._seek_next_stream()
+            elif len(data) < n:
+                n = n - len(data)
+                self._seek_next_stream()
+            else:
+                break
+
+        return b"".join(acc)
+
+    def seekable(self) -> bool:
+        return True
+
+    def writable(self) -> bool:
+        return False
+
+    def readable(self) -> bool:
+        return True
+
+    def seek(self, offset: int, whence: int = io.SEEK_SET) -> int:
+        if whence == io.SEEK_CUR:
+            if offset < 0:
+                raise ValueError("negative offset not supported yet")
+
+            while self._idx < len(self._streams):
+                s = self._streams[self._idx]
+                co = s.tell()
+                eo = s.seek(0, io.SEEK_END)
+                assert co <= eo
+                if offset <= eo - co:
+                    s.seek(co + offset, io.SEEK_SET)
+                    offset = 0
+                    break
+                self._seek_next_stream()
+                offset = offset - (eo - co)
+
+            if 0 < offset:
+                self._offset_after_seek_end += offset
+
+        elif whence == io.SEEK_SET:
+            self._idx = 0
+            self._begin_offset = 0
+            self._offset_after_seek_end = 0
+            self._streams[self._idx].seek(0, io.SEEK_SET)
+            if offset:
+                self.seek(offset, io.SEEK_CUR)
+
+        elif whence == io.SEEK_END:
+            self._idx = 0
+            self._begin_offset = 0
+            self._offset_after_seek_end = 0
+            while self._idx < len(self._streams):
+                self._seek_next_stream()
+            if offset:
+                self.seek(offset, io.SEEK_CUR)
+
+        else:
+            raise IOError("invalid whence")
+
+        return self.tell()
+
+    def tell(self) -> int:
+        if self._idx < len(self._streams):
+            rel_offset = self._streams[self._idx].tell()
+        else:
+            rel_offset = self._offset_after_seek_end
+
+        return self._begin_offset + rel_offset
+
+    def close(self) -> None:
+        for b in self._streams:
+            b.close()
+        return None
+
+
+class SlicedIO(io.IOBase):
+    __slots__ = ("_source", "_begin_offset", "_rel_offset", "_size")
+
+    _source: T.BinaryIO
+    _begin_offset: int
+    _rel_offset: int
+    _size: int
+
+    def __init__(self, source: T.BinaryIO, offset: int, size: int) -> None:
+        assert source.readable(), "source stream must be readable"
+        assert source.seekable(), "source stream must be seekable"
+        self._source = source
+        if offset < 0:
+            raise ValueError(f"negative offset {offset}")
+        self._begin_offset = offset
+        self._rel_offset = 0
+        self._size = size
+
+    def read(self, n: int = -1) -> bytes:
+        if self._rel_offset < self._size:
+            self._source.seek(self._begin_offset + self._rel_offset, io.SEEK_SET)
+            remaining = self._size - self._rel_offset
+            max_read = remaining if n == -1 else min(n, remaining)
+            data = self._source.read(max_read)
+            self._rel_offset += len(data)
+            return data
+        else:
+            return b""
+
+    def seekable(self) -> bool:
+        return True
+
+    def writable(self) -> bool:
+        return False
+
+    def readable(self) -> bool:
+        return True
+
+    def tell(self) -> int:
+        return self._rel_offset
+
+    def seek(self, offset: int, whence: int = io.SEEK_SET) -> int:
+        if whence == io.SEEK_SET:
+            new_offset = offset
+            if new_offset < 0:
+                raise ValueError(f"negative seek value {new_offset}")
+        elif whence == io.SEEK_CUR:
+            new_offset = max(0, self._rel_offset + offset)
+        elif whence == io.SEEK_END:
+            new_offset = max(0, self._size + offset)
+        else:
+            raise IOError("invalid whence")
+        self._rel_offset = new_offset
+        return self._rel_offset
diff --git a/mapillary_tools/geotag/simple_mp4_builder.py b/mapillary_tools/geotag/simple_mp4_builder.py
index acb20a5b..ccf1c8a2 100644
--- a/mapillary_tools/geotag/simple_mp4_builder.py
+++ b/mapillary_tools/geotag/simple_mp4_builder.py
@@ -11,7 +11,7 @@
 
 import construct as C
 
-from . import simple_mp4_parser as parser
+from . import io_utils, simple_mp4_parser as parser
 from .simple_mp4_parser import (
     ChunkLargeOffsetBox,
     ChunkOffsetBox,
@@ -453,7 +453,7 @@ def _filter_moov_children_boxes(
             yield box
 
 
-def _find_movie_timescale(moov_children: T.Sequence[BoxDict]) -> int:
+def find_movie_timescale(moov_children: T.Sequence[BoxDict]) -> int:
     mvhd = _find_box_at_pathx(moov_children, [b"mvhd"])
     return T.cast(T.Dict, mvhd["data"])["timescale"]
 
@@ -467,29 +467,10 @@ def _build_moov_bytes(moov_children: T.Sequence[BoxDict]) -> bytes:
     )
 
 
-class Reader:
-    def read(self):
-        raise NotImplementedError
-
-
-@dataclasses.dataclass
-class SampleReader(Reader):
-    __slots__ = ("fp", "offset", "size")
-
-    fp: T.BinaryIO
-    offset: int
-    size: int
-
-    def read(self):
-        self.fp.seek(self.offset)
-        return self.fp.read(self.size)
-
-
 def transform_mp4(
     src_fp: T.BinaryIO,
-    target_fp: T.BinaryIO,
-    sample_generator: T.Callable[[T.BinaryIO, T.List[BoxDict]], T.Iterator[Reader]],
-):
+    sample_generator: T.Callable[[T.BinaryIO, T.List[BoxDict]], T.Iterator[io.IOBase]],
+) -> io_utils.ChainedIO:
     # extract ftyp
     src_fp.seek(0)
     source_ftyp_box_data = parser.parse_data_firstx(src_fp, [b"ftyp"])
@@ -507,24 +488,28 @@ def transform_mp4(
 
     # extract video samples
     source_samples = list(iterate_samples(moov_children))
-    movie_sample_readers = (
-        SampleReader(src_fp, sample.offset, sample.size) for sample in source_samples
-    )
-
-    sample_readers = itertools.chain(
-        movie_sample_readers, sample_generator(src_fp, moov_children)
-    )
+    movie_sample_readers = [
+        io_utils.SlicedIO(src_fp, sample.offset, sample.size)
+        for sample in source_samples
+    ]
+    sample_readers = list(sample_generator(src_fp, moov_children))
 
     _update_all_trak_tkhd(moov_children)
 
     # moov_boxes should be immutable since here
-    target_fp.write(source_ftyp_data)
-    target_fp.write(rewrite_moov(target_fp.tell(), moov_children))
     mdat_body_size = sum(sample.size for sample in iterate_samples(moov_children))
-    write_mdat(target_fp, mdat_body_size, sample_readers)
+    return io_utils.ChainedIO(
+        [
+            io.BytesIO(source_ftyp_data),
+            io.BytesIO(_rewrite_moov(len(source_ftyp_data), moov_children)),
+            io.BytesIO(_build_mdat_header_bytes(mdat_body_size)),
+            *movie_sample_readers,
+            *sample_readers,
+        ]
+    )
 
 
-def rewrite_moov(moov_offset: int, moov_boxes: T.Sequence[BoxDict]) -> bytes:
+def _rewrite_moov(moov_offset: int, moov_boxes: T.Sequence[BoxDict]) -> bytes:
     # build moov for calculating moov size
     sample_offset = 0
     for box in _filter_trak_boxes(moov_boxes):
@@ -544,10 +529,3 @@ def rewrite_moov(moov_offset: int, moov_boxes: T.Sequence[BoxDict]) -> bytes:
     assert len(moov_data) == moov_data_size, f"{len(moov_data)} != {moov_data_size}"
 
     return moov_data
-
-
-def write_mdat(fp: T.BinaryIO, mdat_body_size: int, sample_readers: T.Iterable[Reader]):
-    mdat_header = _build_mdat_header_bytes(mdat_body_size)
-    fp.write(mdat_header)
-    for reader in sample_readers:
-        fp.write(reader.read())
diff --git a/mapillary_tools/process_import_meta_properties.py b/mapillary_tools/process_import_meta_properties.py
index 0e4dfb1a..4af2716b 100644
--- a/mapillary_tools/process_import_meta_properties.py
+++ b/mapillary_tools/process_import_meta_properties.py
@@ -1,5 +1,5 @@
-import time
 import os
+import time
 import typing as T
 
 from . import exceptions, types, VERSION
diff --git a/tests/integration/test_gopro.py b/tests/integration/test_gopro.py
index 4ea50c3e..ebf33a19 100644
--- a/tests/integration/test_gopro.py
+++ b/tests/integration/test_gopro.py
@@ -1,5 +1,5 @@
-import os
 import json
+import os
 import subprocess
 import typing as T
 from pathlib import Path
diff --git a/tests/unit/test_camm_parser.py b/tests/unit/test_camm_parser.py
index 0b96bb9a..cf8282ef 100644
--- a/tests/unit/test_camm_parser.py
+++ b/tests/unit/test_camm_parser.py
@@ -52,13 +52,11 @@ def build_mp4(points: T.List[geo.Point]) -> T.Optional[T.List[geo.Point]]:
         {"type": b"moov", "data": [mvhd]},
     ]
     src = simple_mp4_builder.QuickBoxStruct32.BoxList.build(empty_mp4)
-    target_fp = io.BytesIO()
-    simple_mp4_builder.transform_mp4(
-        io.BytesIO(src), target_fp, camm_builder.camm_sample_generator2(points)
+    target_fp = simple_mp4_builder.transform_mp4(
+        io.BytesIO(src), camm_builder.camm_sample_generator2(points)
     )
 
-    target_fp.seek(0)
-    return camm_parser.extract_points(target_fp)
+    return camm_parser.extract_points(T.cast(T.BinaryIO, target_fp))
 
 
 def approximate(expected, actual):
diff --git a/tests/unit/test_io_utils.py b/tests/unit/test_io_utils.py
new file mode 100644
index 00000000..1f93c28a
--- /dev/null
+++ b/tests/unit/test_io_utils.py
@@ -0,0 +1,123 @@
+import io
+import random
+
+from mapillary_tools.geotag.io_utils import ChainedIO, SlicedIO
+
+
+def test_chained():
+    data = b"helloworldworldfoobarworld"
+    c = io.BytesIO(data)
+    s = ChainedIO(
+        [
+            io.BytesIO(b"hello"),
+            ChainedIO([io.BytesIO(b"world")]),
+            ChainedIO(
+                [
+                    ChainedIO([io.BytesIO(b""), io.BytesIO(b"")]),
+                    io.BytesIO(b"world"),
+                    io.BytesIO(b"foo"),
+                    ChainedIO([io.BytesIO(b"")]),
+                ]
+            ),
+            ChainedIO([io.BytesIO(b"")]),
+            ChainedIO([io.BytesIO(b"bar")]),
+            ChainedIO(
+                [
+                    SlicedIO(io.BytesIO(data), 5, 5),
+                    ChainedIO([io.BytesIO(b"")]),
+                ]
+            ),
+        ]
+    )
+
+    assert s.seek(0) == 0
+    assert c.seek(0) == 0
+    assert s.read() == c.read()
+
+    assert s.seek(2, io.SEEK_CUR) == len(data) + 2
+    assert c.seek(2, io.SEEK_CUR) == len(data) + 2
+    assert s.read() == c.read()
+
+    assert s.seek(6) == 6
+    assert c.seek(6) == 6
+    assert s.read() == c.read()
+
+    assert s.seek(2, io.SEEK_END) == len(data) + 2
+    assert c.seek(2, io.SEEK_END) == len(data) + 2
+    assert s.read() == c.read()
+
+    assert s.seek(0) == 0
+    assert c.seek(0) == 0
+    assert s.read(1) == b"h"
+    assert s.read(1000) == data[1:]
+    assert s.read() == b""
+    assert s.read(1) == b""
+
+    assert s.seek(0, io.SEEK_END) == len(data)
+    assert c.seek(0, io.SEEK_END) == len(data)
+
+    c.seek(0)
+    s.seek(0)
+    for _ in range(10000):
+        whence = random.choice([io.SEEK_SET, io.SEEK_CUR, io.SEEK_END])
+        offset = random.randint(0, 30)
+        assert s.tell() == c.tell()
+        thrown_x = None
+        try:
+            x = s.seek(offset, whence)
+        except ValueError as ex:
+            thrown_x = ex
+        thrown_y = None
+        try:
+            y = c.seek(offset, whence)
+        except ValueError as ex:
+            thrown_y = ex
+        assert (thrown_x is not None and thrown_y is not None) or (
+            thrown_x is None and thrown_y is None
+        ), (thrown_x, thrown_y, whence, offset)
+        if not thrown_x:
+            assert (
+                x == y
+            ), f"whence={whence} offset={offset} x={x} y={y} {s.tell()} {c.tell()}"
+
+        n = random.randint(-1, 20)
+        assert s.read(n) == c.read(n), f"n={n}"
+        assert s.tell() == c.tell()
+
+
+def test_sliced():
+    s = io.BytesIO(b"helloworldfoo")
+    sliced = SlicedIO(s, 5, 5)
+    c = io.BytesIO(b"world")
+
+    for _ in range(10000):
+        whence = random.choice([io.SEEK_SET, io.SEEK_CUR, io.SEEK_END])
+        offset = random.randint(-10, 10)
+        thrown_x = None
+        try:
+            x = sliced.seek(offset, whence)
+        except ValueError as ex:
+            thrown_x = ex
+        thrown_y = None
+        try:
+            y = c.seek(offset, whence)
+        except ValueError as ex:
+            thrown_y = ex
+        assert (thrown_x is not None and thrown_y is not None) or (
+            thrown_x is None and thrown_y is None
+        ), (thrown_x, thrown_y, whence, offset)
+        if not thrown_x:
+            assert x == y
+
+        n = random.randint(-1, 20)
+        assert sliced.read(n) == c.read(n)
+        assert sliced.tell() == c.tell()
+
+
+def test_truncate():
+    c = io.BytesIO(b"helloworld")
+    c.truncate(3)
+    assert c.read() == b"hel"
+    s = SlicedIO(c, 1, 5)
+    assert s.read() == b"el"
+    assert s.read() == b""