From 099e54cc1b20e5d8ba0b14ba7dd8c989dfa9c7fa Mon Sep 17 00:00:00 2001
From: Boris Sekachev <boris.sekachev@yandex.ru>
Date: Tue, 5 Sep 2023 16:25:13 +0300
Subject: [PATCH] Fixed error for non-jpeg compressed images in original chunks
 (#6789)

Resolved #6598
---
 CHANGELOG.md                         |  1 +
 cvat/apps/engine/media_extractors.py | 22 +++++++++++++++++++---
 2 files changed, 20 insertions(+), 3 deletions(-)

diff --git a/CHANGELOG.md b/CHANGELOG.md
index 86071589ce77..d3c79bb34cba 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -23,6 +23,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 ### Fixed
 - Zooming canvas when scrooling comments list in an issue (<https://github.com/opencv/cvat/pull/6758>)
 - Issues can be created many times when initial submit (<https://github.com/opencv/cvat/pull/6758>)
+- Running deep learning models on non-jpeg compressed tif images (<https://github.com/opencv/cvat/pull/6789>)
 - Paddings on tasks/projects/models pages (<https://github.com/opencv/cvat/pull/6778>)
 - Memory leak in the logging system (<https://github.com/opencv/cvat/pull/6804>)
 
diff --git a/cvat/apps/engine/media_extractors.py b/cvat/apps/engine/media_extractors.py
index 65ab80ae4745..c81efc82b145 100644
--- a/cvat/apps/engine/media_extractors.py
+++ b/cvat/apps/engine/media_extractors.py
@@ -80,6 +80,9 @@ def image_size_within_orientation(img: Image):
         return img.height, img.width
     return img.width, img.height
 
+def has_exif_rotation(img: Image):
+    return img.getexif().get(ORIENTATION_EXIF_TAG, ORIENTATION.NORMAL_HORIZONTAL) != ORIENTATION.NORMAL_HORIZONTAL
+
 def rotate_within_exif(img: Image):
     orientation = img.getexif().get(ORIENTATION_EXIF_TAG,  ORIENTATION.NORMAL_HORIZONTAL)
     if orientation in [ORIENTATION.NORMAL_180_ROTATED, ORIENTATION.MIRROR_VERTICAL]:
@@ -661,12 +664,25 @@ def save_as_chunk(self, images, chunk_path):
                 ext = os.path.splitext(path)[1].replace('.', '')
                 output = io.BytesIO()
                 if self._dimension == DimensionType.DIM_2D:
-                    pil_image = rotate_within_exif(Image.open(image))
-                    pil_image.save(output, format=pil_image.format if pil_image.format else self.IMAGE_EXT, quality=100, subsampling=0)
+                    pil_image = Image.open(image)
+                    if has_exif_rotation(pil_image):
+                        rot_image = rotate_within_exif(pil_image)
+                        if rot_image.format == 'TIFF':
+                            # https://pillow.readthedocs.io/en/stable/handbook/image-file-formats.html
+                            # use loseless lzw compression for tiff images
+                            rot_image.save(output, format='TIFF', compression='tiff_lzw')
+                        else:
+                            rot_image.save(output, format=rot_image.format if rot_image.format else self.IMAGE_EXT, quality=100, subsampling=0)
+                    else:
+                        output = image
                 else:
                     output, ext = self._write_pcd_file(image)[0:2]
                 arcname = '{:06d}.{}'.format(idx, ext)
-                zip_chunk.writestr(arcname, output.getvalue())
+
+                if isinstance(output, io.BytesIO):
+                    zip_chunk.writestr(arcname, output.getvalue())
+                else:
+                    zip_chunk.write(filename=output, arcname=arcname)
         # return empty list because ZipChunkWriter write files as is
         # and does not decode it to know img size.
         return []