added pillow backend in loading (#333)

* added pillow backend in loading

* reformat

* remove auto fill for RGBA

* change according to comments

* change according to comments

* reformat

docs/make.bat

@@ -1,35 +1,35 @@
-@ECHO OFF
-
-pushd %~dp0
-
-REM Command file for Sphinx documentation
-
-if "%SPHINXBUILD%" == "" (
-	set SPHINXBUILD=sphinx-build
-)
-set SOURCEDIR=.
-set BUILDDIR=_build
-
-if "%1" == "" goto help
-
-%SPHINXBUILD% >NUL 2>NUL
-if errorlevel 9009 (
-	echo.
-	echo.The 'sphinx-build' command was not found. Make sure you have Sphinx
-	echo.installed, then set the SPHINXBUILD environment variable to point
-	echo.to the full path of the 'sphinx-build' executable. Alternatively you
-	echo.may add the Sphinx directory to PATH.
-	echo.
-	echo.If you don't have Sphinx installed, grab it from
-	echo.http://sphinx-doc.org/
-	exit /b 1
-)
-
-%SPHINXBUILD% -M %1 %SOURCEDIR% %BUILDDIR% %SPHINXOPTS%
-goto end
-
-:help
-%SPHINXBUILD% -M help %SOURCEDIR% %BUILDDIR% %SPHINXOPTS%
-
-:end
-popd
+@ECHO OFF
+
+pushd %~dp0
+
+REM Command file for Sphinx documentation
+
+if "%SPHINXBUILD%" == "" (
+	set SPHINXBUILD=sphinx-build
+)
+set SOURCEDIR=.
+set BUILDDIR=_build
+
+if "%1" == "" goto help
+
+%SPHINXBUILD% >NUL 2>NUL
+if errorlevel 9009 (
+	echo.
+	echo.The 'sphinx-build' command was not found. Make sure you have Sphinx
+	echo.installed, then set the SPHINXBUILD environment variable to point
+	echo.to the full path of the 'sphinx-build' executable. Alternatively you
+	echo.may add the Sphinx directory to PATH.
+	echo.
+	echo.If you don't have Sphinx installed, grab it from
+	echo.http://sphinx-doc.org/
+	exit /b 1
+)
+
+%SPHINXBUILD% -M %1 %SOURCEDIR% %BUILDDIR% %SPHINXOPTS%
+goto end
+
+:help
+%SPHINXBUILD% -M help %SOURCEDIR% %BUILDDIR% %SPHINXOPTS%
+
+:end
+popd

mmcv/image/io.py

@@ -1,4 +1,5 @@
 # Copyright (c) Open-MMLab. All rights reserved.
+import io
 import os.path as osp
 from pathlib import Path
 
@@ -13,8 +14,13 @@ try:
 except ImportError:
     TJCS_RGB = TJPF_GRAY = TJPF_BGR = TurboJPEG = None
 
+try:
+    from PIL import Image
+except ImportError:
+    Image = None
+
 jpeg = None
-supported_backends = ['cv2', 'turbojpeg']
+supported_backends = ['cv2', 'turbojpeg', 'pillow']
 
 imread_flags = {
     'color': IMREAD_COLOR,
@@ -42,6 +48,9 @@ def use_backend(backend):
         global jpeg
         if jpeg is None:
             jpeg = TurboJPEG()
+    elif imread_backend == 'pillow':
+        if Image is None:
+            raise ImportError('`Pillow` is not installed')
 
 
 def _jpegflag(flag='color', channel_order='bgr'):
@@ -60,6 +69,57 @@ def _jpegflag(flag='color', channel_order='bgr'):
         raise ValueError('flag must be "color" or "grayscale"')
 
 
+def _pillow2array(img, flag='color', channel_order='bgr'):
+    """Convert a pillow image to numpy array
+
+    Args:
+        img (:obj:`PIL.Image.Image`): The image loaded using PIL
+        flag (str): Flags specifying the color type of a loaded image,
+            candidates are 'color', 'grayscale' and 'unchanged'.
+            Default to 'color'.
+        channel_order (str): The channel order of the output image array,
+            candidates are 'bgr' and 'rgb'. Default to 'bgr'.
+
+    Returns:
+        np.ndarray: The converted numpy array
+    """
+    channel_order = channel_order.lower()
+    if channel_order not in ['rgb', 'bgr']:
+        raise ValueError('channel order must be either "rgb" or "bgr"')
+
+    if flag == 'unchanged':
+        array = np.array(img)
+        if array.ndim >= 3 and array.shape[2] >= 3:  # color image
+            array[:, :, :3] = array[:, :, (2, 1, 0)]  # RGB to BGR
+    else:
+        # If the image mode is not 'RGB', convert it to 'RGB' first.
+        if img.mode != 'RGB':
+            if img.mode != 'LA':
+                # Most formats except 'LA' can be directly converted to RGB
+                img = img.convert('RGB')
+            else:
+                # When the mode is 'LA', the default conversion will fill in
+                #  the canvas with black, which sometimes shadows black objects
+                #  in the foreground.
+                #
+                # Therefore, a random color (124, 117, 104) is used for canvas
+                img_rgba = img.convert('RGBA')
+                img = Image.new('RGB', img_rgba.size, (124, 117, 104))
+                img.paste(img_rgba, mask=img_rgba.split()[3])  # 3 is alpha
+        if flag == 'color':
+            array = np.array(img)
+            if channel_order != 'rgb':
+                array = array[:, :, ::-1]  # RGB to BGR
+        elif flag == 'grayscale':
+            img = img.convert('L')
+            array = np.array(img)
+        else:
+            raise ValueError(
+                'flag must be "color", "grayscale" or "unchanged", '
+                f'but got {flag}')
+    return array
+
+
 def imread(img_or_path, flag='color', channel_order='bgr'):
     """Read an image.
 
@@ -90,6 +150,10 @@ def imread(img_or_path, flag='color', channel_order='bgr'):
                 if img.shape[-1] == 1:
                     img = img[:, :, 0]
             return img
+        elif imread_backend == 'pillow':
+            img = Image.open(img_or_path)
+            img = _pillow2array(img, flag, channel_order)
+            return img
         else:
             flag = imread_flags[flag] if is_str(flag) else flag
             img = cv2.imread(img_or_path, flag)
@@ -116,6 +180,11 @@ def imfrombytes(content, flag='color', channel_order='bgr'):
         if img.shape[-1] == 1:
             img = img[:, :, 0]
         return img
+    elif imread_backend == 'pillow':
+        buff = io.BytesIO(content)
+        img = Image.open(buff)
+        img = _pillow2array(img, flag, channel_order)
+        return img
     else:
         img_np = np.frombuffer(content, np.uint8)
         flag = imread_flags[flag] if is_str(flag) else flag

mmcv/model_zoo/open_mmlab.json

@@ -32,4 +32,4 @@
   "regnetx_6.4gf": "https://open-mmlab.s3.ap-northeast-2.amazonaws.com/pretrain/third_party/regnetx_6.4gf-006af45d.pth",
   "regnetx_8.0gf": "https://open-mmlab.s3.ap-northeast-2.amazonaws.com/pretrain/third_party/regnetx_8.0gf-3c68abe7.pth",
   "regnetx_12gf": "https://open-mmlab.s3.ap-northeast-2.amazonaws.com/pretrain/third_party/regnetx_12gf-4c2a3350.pth"
-}
+}

tests/test_image/test_io.py

@@ -8,7 +8,7 @@ from unittest.mock import patch
 import cv2
 import numpy as np
 import pytest
-from numpy.testing import assert_array_equal
+from numpy.testing import assert_allclose, assert_array_equal
 
 import mmcv
 
@@ -65,6 +65,51 @@ class TestIO:
         with pytest.raises(TypeError):
             mmcv.imread(1)
 
+        # backend pillow
+        mmcv.use_backend('pillow')
+        self.gray_alpha_img_path = osp.join(self.data_dir, 'gray_alpha.png')
+        self.palette_img_path = osp.join(self.data_dir, 'palette.gif')
+        img_pil_grayscale1 = mmcv.imread(self.img_path, 'grayscale')
+        assert img_pil_grayscale1.shape == (300, 400)
+        img_pil_gray_alpha = mmcv.imread(self.gray_alpha_img_path, 'grayscale')
+        assert img_pil_gray_alpha.shape == (400, 500)
+        mean = img_pil_gray_alpha[300:, 400:].mean()
+        assert_allclose(img_pil_gray_alpha[300:, 400:] - mean, 0)
+        img_pil_gray_alpha = mmcv.imread(self.gray_alpha_img_path)
+        mean = img_pil_gray_alpha[300:, 400:].mean(axis=(0, 1))
+        assert_allclose(img_pil_gray_alpha[300:, 400:] - mean, 0)
+        assert img_pil_gray_alpha.shape == (400, 500, 3)
+        img_pil_gray_alpha = mmcv.imread(self.gray_alpha_img_path, 'unchanged')
+        assert img_pil_gray_alpha.shape == (400, 500, 2)
+        img_pil_palette = mmcv.imread(self.palette_img_path, 'grayscale')
+        assert img_pil_palette.shape == (300, 400)
+        img_pil_palette = mmcv.imread(self.palette_img_path)
+        assert img_pil_palette.shape == (300, 400, 3)
+        img_pil_palette = mmcv.imread(self.palette_img_path, 'unchanged')
+        assert img_pil_palette.shape == (300, 400)
+        img_pil_grayscale2 = mmcv.imread(self.gray_img_path)
+        assert img_pil_grayscale2.shape == (300, 400, 3)
+        img_pil_unchanged = mmcv.imread(self.gray_img_path, 'unchanged')
+        assert img_pil_unchanged.shape == (300, 400)
+        img_pil_unchanged = mmcv.imread(img_pil_unchanged)
+        assert_array_equal(img_pil_unchanged, mmcv.imread(img_pil_unchanged))
+
+        img_pil_color_bgr = mmcv.imread(self.img_path_obj)
+        assert img_pil_color_bgr.shape == (300, 400, 3)
+        img_pil_color_rgb = mmcv.imread(self.img_path_obj, channel_order='rgb')
+        assert img_pil_color_rgb.shape == (300, 400, 3)
+        assert (img_pil_color_rgb == img_cv2_color_rgb).sum() / float(
+            img_cv2_color_rgb.size) > 0.5
+        assert_array_equal(img_pil_color_rgb[:, :, ::-1], img_pil_color_bgr)
+        img_pil_grayscale1 = mmcv.imread(self.img_path_obj, 'grayscale')
+        assert img_pil_grayscale1.shape == (300, 400)
+        img_pil_grayscale2 = mmcv.imread(self.gray_img_path_obj)
+        assert img_pil_grayscale2.shape == (300, 400, 3)
+        img_pil_unchanged = mmcv.imread(self.gray_img_path_obj, 'unchanged')
+        assert img_pil_unchanged.shape == (300, 400)
+        with pytest.raises(TypeError):
+            mmcv.imread(1)
+
         # backend turbojpeg
         mmcv.use_backend('turbojpeg')
 
@@ -137,6 +182,15 @@ class TestIO:
         gray_img_dim3_cv2 = mmcv.imfrombytes(img_bytes, flag='grayscale')
         assert gray_img_dim3_cv2.shape == (300, 400)
 
+        # backend pillow, channel order: bgr
+        mmcv.use_backend('pillow')
+        with open(self.img_path, 'rb') as f:
+            img_bytes = f.read()
+        img_pillow = mmcv.imfrombytes(img_bytes)
+        assert img_pillow.shape == (300, 400, 3)
+        # Pillow and opencv decoding may not be the same
+        assert (img_cv2 == img_pillow).sum() / float(img_cv2.size) > 0.5
+
         # backend turbojpeg, channel order: bgr
         mmcv.use_backend('turbojpeg')
         with open(self.img_path, 'rb') as f: