Loading 16-bit tiff images (cvat-ai#5005)

Pull Request regarding Issue cvat-ai#2987 

PIL.Image conversion from I;16 to L or RGB are unsuccessful as for now.
See the corresponding Issue in the Pillow GitHub (Opened 2018, so no
changes to be expected)
python-pillow/Pillow#3011

The proposed changes at least fix this issue for the mode 'I;16' and
delivers a possible solution for other modes (eg. I;16B/L/N).

This results in a correct calculation of the preview thumbnail and the
actual image, the annotation will be performed on.

We have used this solution on our own dataset and created annotations
accordingly.

CHANGELOG.md

@@ -83,6 +83,7 @@ from online detectors & interactors) (<https://github.com/opencv/cvat/pull/4543>
 non-ascii paths while adding files from "Connected file share" (issue #4428)
 - Removed unnecessary volumes defined in docker-compose.serverless.yml
 (<https://github.com/openvinotoolkit/cvat/pull/4659>)
+- Added support for Image files that use the PIL.Image.mode 'I;16'
 - Project import/export with skeletons (<https://github.com/opencv/cvat/pull/4867>,
   <https://github.com/opencv/cvat/pull/5004>)
 - Shape color is not changed on canvas after changing a label (<https://github.com/opencv/cvat/pull/5045>)

cvat/apps/engine/media_extractors.py

@@ -17,7 +17,7 @@
 import numpy as np
 from natsort import os_sorted
 from pyunpack import Archive
-from PIL import Image, ImageFile
+from PIL import Image, ImageFile, ImageOps
 from random import shuffle
 from cvat.apps.engine.utils import rotate_image
 from cvat.apps.engine.models import DimensionType, SortingMethod
@@ -127,9 +127,25 @@ def _get_preview(obj):
         else:
             preview = obj
         preview = rotate_within_exif(preview)
+        # TODO - Check if the other formats work. I'm only interested in I;16 for now. Sorry @:-|
+        # Summary:
+        # Images in the Format I;16 definitely don't work. Most likely I;16B/L/N won't work as well.
+        # Simple Conversion from I;16 to I/RGB/L doesn't work as well.
+        #   Including any Intermediate Conversions doesn't work either. (eg. I;16 to I to L)
+        # Seems like an internal Bug of PIL
+        #     See Issue for further details: https://github.com/python-pillow/Pillow/issues/3011
+        #     Issue was opened 2018, so don't expect any changes soon and work with manual conversions.
+        mode: str = preview.mode
+        if mode == "I;16":
+            preview = np.array(preview, dtype=np.uint16) # 'I;16' := Unsigned Integer 16, Grayscale
+            image = image - image.min()                  # In case the used range lies in [a, 2^16] with a > 0
+            preview = preview / preview.max() * 255      # Downscale into real numbers of range [0, 255]
+            preview = preview.astype(np.uint8)           # Floor to integers of range [0, 255]
+            preview = Image.fromarray(preview, mode="L") # 'L' := Unsigned Integer 8, Grayscale
+            preview = ImageOps.equalize(preview)         # The Images need equalization. High resolution with 16-bit but only small range that actually contains information
         preview.thumbnail(PREVIEW_SIZE)
 
-        return preview.convert('RGB')
+        return preview
 
     @abstractmethod
     def get_image_size(self, i):
@@ -589,6 +605,23 @@ def _compress_image(image_path, quality):
             im_data = np.array(image)
             im_data = im_data * (2**8 / im_data.max())
             image = Image.fromarray(im_data.astype(np.int32))
+
+        # TODO - Check if the other formats work. I'm only interested in I;16 for now. Sorry @:-|
+        # Summary:
+        # Images in the Format I;16 definitely don't work. Most likely I;16B/L/N won't work as well.
+        # Simple Conversion from I;16 to I/RGB/L doesn't work as well.
+        #   Including any Intermediate Conversions doesn't work either. (eg. I;16 to I to L)
+        # Seems like an internal Bug of PIL
+        #     See Issue for further details: https://github.com/python-pillow/Pillow/issues/3011
+        #     Issue was opened 2018, so don't expect any changes soon and work with manual conversions.
+        if image.mode == "I;16":
+            image = np.array(image, dtype=np.uint16) # 'I;16' := Unsigned Integer 16, Grayscale
+            image = image - image.min()              # In case the used range lies in [a, 2^16] with a > 0
+            image = image / image.max() * 255        # Downscale into real numbers of range [0, 255]
+            image = image.astype(np.uint8)           # Floor to integers of range [0, 255]
+            image = Image.fromarray(image, mode="L") # 'L' := Unsigned Integer 8, Grayscale
+            image = ImageOps.equalize(image)         # The Images need equalization. High resolution with 16-bit but only small range that actually contains information
+
         converted_image = image.convert('RGB')
         image.close()
         buf = io.BytesIO()