From de10689433c292ee5eac5bb20a327b9c1d8b71f9 Mon Sep 17 00:00:00 2001
From: Hye Soo Yang <hyey@google.com>
Date: Mon, 19 Oct 2020 04:40:07 -0700
Subject: [PATCH] Fix DecodeImage issue where decoding animated GIF fails if
first frame does not fill the canvas.
PiperOrigin-RevId: 337830066
Change-Id: Iecdf4cb7f6815f65e7abceb85b61e8d85d80fa9e
---
tensorflow/core/BUILD | 2 +
.../base_api/api_def_DecodeImage.pbtxt | 6 +++
tensorflow/core/lib/gif/gif_io.cc | 41 +++++++++++-------
.../core/lib/gif/testdata/red_black.gif | Bin 0 -> 88 bytes
tensorflow/core/lib/gif/testdata/squares.gif | Bin 0 -> 103 bytes
tensorflow/python/ops/image_ops_impl.py | 6 +++
tensorflow/python/ops/image_ops_test.py | 28 ++++++++++++
7 files changed, 67 insertions(+), 16 deletions(-)
create mode 100644 tensorflow/core/lib/gif/testdata/red_black.gif
create mode 100644 tensorflow/core/lib/gif/testdata/squares.gif
diff --git a/tensorflow/core/BUILD b/tensorflow/core/BUILD
index adc59c67dce..943ead9d451 100644
--- a/tensorflow/core/BUILD
+++ b/tensorflow/core/BUILD
@@ -2128,6 +2128,8 @@ filegroup(
# GIF data
"lib/gif/testdata/lena.gif",
"lib/gif/testdata/scan.gif",
+ "lib/gif/testdata/red_black.gif",
+ "lib/gif/testdata/squares.gif",
# GIF data with optimization
"lib/gif/testdata/optimized.gif",
# BMP data
diff --git a/tensorflow/core/api_def/base_api/api_def_DecodeImage.pbtxt b/tensorflow/core/api_def/base_api/api_def_DecodeImage.pbtxt
index c534425eb24..7174c8d3daf 100644
--- a/tensorflow/core/api_def/base_api/api_def_DecodeImage.pbtxt
+++ b/tensorflow/core/api_def/base_api/api_def_DecodeImage.pbtxt
@@ -47,5 +47,11 @@ constructing your graph if you are intermixing GIF files with BMP, JPEG, and/or
PNG files. Alternately, set the expand_animations argument of this function to
False, in which case the op will return 3-dimensional tensors and will truncate
animated GIF files to the first frame.
+
+*NOTE*: If the first frame of an animated GIF does not occupy the entire
+canvas (maximum frame width x maximum frame height), then it fills the
+unoccupied areas (in the first frame) with zeros (black). For frames after the
+first frame that does not occupy the entire canvas, it uses the previous
+frame to fill the unoccupied areas.
END
}
diff --git a/tensorflow/core/lib/gif/gif_io.cc b/tensorflow/core/lib/gif/gif_io.cc
index 659513d05ed..5fb47043654 100644
--- a/tensorflow/core/lib/gif/gif_io.cc
+++ b/tensorflow/core/lib/gif/gif_io.cc
@@ -85,7 +85,6 @@ uint8* Decode(const void* srcdata, int datasize,
}
int target_num_frames = gif_file->ImageCount;
- if (!expand_animations) target_num_frames = 1;
// Don't request more memory than needed for each frame, preventing OOM
int max_frame_width = 0;
@@ -101,6 +100,7 @@ uint8* Decode(const void* srcdata, int datasize,
const int width = max_frame_width;
const int height = max_frame_height;
const int channel = 3;
+ if (!expand_animations) target_num_frames = 1;
uint8* const dstdata =
allocate_output(target_num_frames, width, height, channel);
@@ -118,27 +118,36 @@ uint8* Decode(const void* srcdata, int datasize,
if (img_desc->Left != 0 || img_desc->Top != 0 || img_desc->Width != width ||
img_desc->Height != height) {
- // If the first frame does not fill the entire canvas then return error.
+ // If the first frame does not fill the entire canvas then fill the
+ // unoccupied canvas with zeros (black).
if (k == 0) {
- *error_string = "the first frame does not fill the canvas";
- return nullptr;
+ for (int i = 0; i < height; ++i) {
+ uint8* p_dst = this_dst + i * width * channel;
+ for (int j = 0; j < width; ++j) {
+ p_dst[j * channel + 0] = 0;
+ p_dst[j * channel + 1] = 0;
+ p_dst[j * channel + 2] = 0;
+ }
+ }
+ } else {
+ // Otherwise previous frame will be reused to fill the unoccupied
+ // canvas.
+ uint8* last_dst = dstdata + (k - 1) * width * channel * height;
+ for (int i = 0; i < height; ++i) {
+ uint8* p_dst = this_dst + i * width * channel;
+ uint8* l_dst = last_dst + i * width * channel;
+ for (int j = 0; j < width; ++j) {
+ p_dst[j * channel + 0] = l_dst[j * channel + 0];
+ p_dst[j * channel + 1] = l_dst[j * channel + 1];
+ p_dst[j * channel + 2] = l_dst[j * channel + 2];
+ }
+ }
}
- // Otherwise previous frame will be reused to fill the unoccupied canvas.
+
imgLeft = std::max(imgLeft, 0);
imgTop = std::max(imgTop, 0);
imgRight = std::min(imgRight, width);
imgBottom = std::min(imgBottom, height);
-
- uint8* last_dst = dstdata + (k - 1) * width * channel * height;
- for (int i = 0; i < height; ++i) {
- uint8* p_dst = this_dst + i * width * channel;
- uint8* l_dst = last_dst + i * width * channel;
- for (int j = 0; j < width; ++j) {
- p_dst[j * channel + 0] = l_dst[j * channel + 0];
- p_dst[j * channel + 1] = l_dst[j * channel + 1];
- p_dst[j * channel + 2] = l_dst[j * channel + 2];
- }
- }
}
ColorMapObject* color_map = this_image->ImageDesc.ColorMap
diff --git a/tensorflow/core/lib/gif/testdata/red_black.gif b/tensorflow/core/lib/gif/testdata/red_black.gif
new file mode 100644
index 0000000000000000000000000000000000000000..d32ddd3547d4c7601fa02acc725db8435e4a3fa4
GIT binary patch
literal 88
zcmZ?wbhEHb6krfw_`m=H$22q)|8x7fh6Fo12DlpO889;fMHPRtaELJcXV3vD0V!r+
ll4<E*d3r75;+jpDr1bfu%4g?9Y7{N`;d}1XHV;MyYXFAd7x@4H
literal 0
HcmV?d00001
diff --git a/tensorflow/core/lib/gif/testdata/squares.gif b/tensorflow/core/lib/gif/testdata/squares.gif
new file mode 100644
index 0000000000000000000000000000000000000000..159f86355a8fd7af93a7774531c10134117337f1
GIT binary patch
literal 103
zcmZ?wbhEHb6krfw_`tw$OhZF)+GoZ8+<vYh!Oo5Wu10zW%!~{S42nNlIE)znGw3ig
sFf(ugAp;XfOaIE#Z~2q85t2ZKKyw6ul6*a2$;ETF-0I$J$I4(004eDk3IG5A
literal 0
HcmV?d00001
diff --git a/tensorflow/python/ops/image_ops_impl.py b/tensorflow/python/ops/image_ops_impl.py
index 8a767597f00..75e66d8f513 100644
--- a/tensorflow/python/ops/image_ops_impl.py
+++ b/tensorflow/python/ops/image_ops_impl.py
@@ -2964,6 +2964,12 @@ def decode_image(contents,
function to `False`, in which case the op will return 3-dimensional tensors
and will truncate animated GIF files to the first frame.
+ NOTE: If the first frame of an animated GIF does not occupy the entire
+ canvas (maximum frame width x maximum frame height), then it fills the
+ unoccupied areas (in the first frame) with zeros (black). For frames after the
+ first frame that does not occupy the entire canvas, it uses the previous
+ frame to fill the unoccupied areas.
+
Args:
contents: 0-D `string`. The encoded image bytes.
channels: An optional `int`. Defaults to `0`. Number of color channels for
diff --git a/tensorflow/python/ops/image_ops_test.py b/tensorflow/python/ops/image_ops_test.py
index 7f4a093ca40..a7c964d8a30 100644
--- a/tensorflow/python/ops/image_ops_test.py
+++ b/tensorflow/python/ops/image_ops_test.py
@@ -5756,6 +5756,34 @@ class DecodeImageTest(test_util.TensorFlowTestCase, parameterized.TestCase):
img = image_ops.decode_and_crop_jpeg(img_bytes, [1, 1, 2, 2])
self.evaluate(img)
+ def testGifFramesWithDiffSize(self):
+ """Test decoding an animated GIF.
+
+ This test verifies that `decode_image` op can decode animated GIFs whose
+ first frame does not fill the canvas. The unoccupied areas should be filled
+ with zeros (black).
+
+ `squares.gif` is animated with two images of different sizes. It
+ alternates between a smaller image of size 10 x 10 and a larger image of
+ size 16 x 16. Because it starts animating with the smaller image, the first
+ frame does not fill the canvas. (Canvas size is equal to max frame width x
+ max frame height.)
+
+ `red_black.gif` has just a single image in a GIF format. It is the same
+ image as the smaller image (size 10 x 10) of the two images in
+ `squares.gif`. The only difference is that its background (canvas - smaller
+ image) is pre-filled with zeros (black); it is the groundtruth.
+ """
+ base = "tensorflow/core/lib/gif/testdata"
+ gif_bytes0 = io_ops.read_file(os.path.join(base, "squares.gif"))
+ image0 = image_ops.decode_image(gif_bytes0, dtype=dtypes.float32,
+ expand_animations=False)
+ gif_bytes1 = io_ops.read_file(os.path.join(base, "red_black.gif"))
+ image1 = image_ops.decode_image(gif_bytes1, dtype=dtypes.float32)
+ image1_0 = array_ops.gather(image1, 0)
+ image0, image1_0 = self.evaluate([image0, image1_0])
+ self.assertAllEqual(image0, image1_0)
+
if __name__ == "__main__":
googletest.main()