Ignore empty lines in YOLO annotations (cvat-ai#221)

* Ignore empty lines in yolo annotations

* Add type hints for image class, catch image opening errors in image.size

* update changelog

CHANGELOG.md

@@ -27,6 +27,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - Image extension in CVAT format export (<https://github.com/openvinotoolkit/datumaro/pull/214>)
 - Added a label "face" for bounding boxes in Wider Face (<https://github.com/openvinotoolkit/datumaro/pull/215>)
 - Allowed adding "difficult", "truncated", "occluded" attributes when converting to Pascal VOC if these attributes are not present (<https://github.com/openvinotoolkit/datumaro/pull/216>)
+- Empty lines in YOLO annotations are ignored (<https://github.com/openvinotoolkit/datumaro/pull/221>)
 
 ### Security
 -

datumaro/components/operations.py

@@ -957,6 +957,11 @@ def mean_std(dataset):
     var = lambda i, s: s[i][1]
 
     for i, item in enumerate(dataset):
+        size = item.image.size
+        if size is None:
+            log.warning("Item %s: can't detect image size, "
+                "the image will be skipped from pixel statistics", item.id)
+            continue
         counts[i] = np.prod(item.image.size)
 
         image = item.image.data

datumaro/plugins/yolo_format/extractor.py

@@ -92,7 +92,7 @@ def __init__(self, config_path, image_info=None):
             with open(list_path, 'r', encoding='utf-8') as f:
                 subset.items = OrderedDict(
                     (self.name_from_path(p), self.localize_path(p))
-                    for p in f
+                    for p in f if p.strip()
                 )
             subsets[subset_name] = subset
 
@@ -176,7 +176,9 @@ def _load_categories(names_path):
 
         with open(names_path, 'r', encoding='utf-8') as f:
             for label in f:
-                label_categories.add(label.strip())
+                label = label.strip()
+                if label:
+                    label_categories.add(label)
 
         return label_categories
 

datumaro/util/image.py

@@ -7,19 +7,21 @@
 
 from enum import Enum
 from io import BytesIO
-from typing import Iterator, Iterable, Union
+from typing import Any, Callable, Iterator, Iterable, Optional, Tuple, Union
 import numpy as np
 import os
 import os.path as osp
 
 _IMAGE_BACKENDS = Enum('_IMAGE_BACKENDS', ['cv2', 'PIL'])
 _IMAGE_BACKEND = None
+_image_loading_errors = (FileNotFoundError, )
 try:
     import cv2
     _IMAGE_BACKEND = _IMAGE_BACKENDS.cv2
 except ImportError:
     import PIL
     _IMAGE_BACKEND = _IMAGE_BACKENDS.PIL
+    _image_loading_errors = (*_image_loading_errors, PIL.UnidentifiedImageError)
 
 from datumaro.util.image_cache import ImageCache as _ImageCache
 from datumaro.util.os_util import walk
@@ -33,6 +35,8 @@ def load_image(path, dtype=np.float32):
     if _IMAGE_BACKEND == _IMAGE_BACKENDS.cv2:
         import cv2
         image = cv2.imread(path, cv2.IMREAD_UNCHANGED)
+        if image is None:
+            raise FileNotFoundError("Can't open image: %s" % path)
         image = image.astype(dtype)
     elif _IMAGE_BACKEND == _IMAGE_BACKENDS.PIL:
         from PIL import Image
@@ -43,8 +47,6 @@ def load_image(path, dtype=np.float32):
     else:
         raise NotImplementedError()
 
-    if image is None:
-        raise ValueError("Can't open image '%s'" % path)
     assert len(image.shape) in {2, 3}
     if len(image.shape) == 3:
         assert image.shape[2] in {3, 4}
@@ -227,15 +229,16 @@ def __hash__(self):
         return hash((id(self), self.path, self.loader))
 
 class Image:
-    def __init__(self, data=None, path=None, loader=None, cache=None,
-            size=None):
-        assert size is None or len(size) == 2
+    def __init__(self, data: Union[None, Callable, np.ndarray] = None,
+            path: Optional[str] = None, loader: Optional[Callable] = None,
+            size: Optional[Tuple[int, int]] = None, cache: Any = None):
+        assert size is None or len(size) == 2, size
         if size is not None:
             assert len(size) == 2 and 0 < size[0] and 0 < size[1], size
             size = tuple(size)
         self._size = size # (H, W)
 
-        assert path is None or isinstance(path, str)
+        assert path is None or isinstance(path, str), path
         if path is None:
             path = ''
         elif path:
@@ -254,15 +257,15 @@ def __init__(self, data=None, path=None, loader=None, cache=None,
             self._size = data.shape[:2]
 
     @property
-    def path(self):
+    def path(self) -> str:
         return self._path
 
     @property
-    def ext(self):
+    def ext(self) -> str:
         return osp.splitext(osp.basename(self.path))[1]
 
     @property
-    def data(self):
+    def data(self) -> np.ndarray:
         if callable(self._data):
             data = self._data()
         else:
@@ -273,17 +276,20 @@ def data(self):
         return data
 
     @property
-    def has_data(self):
+    def has_data(self) -> bool:
         return self._data is not None
 
     @property
-    def has_size(self):
+    def has_size(self) -> bool:
         return self._size is not None or isinstance(self._data, np.ndarray)
 
     @property
-    def size(self):
+    def size(self) -> Optional[Tuple[int, int]]:
         if self._size is None:
-            data = self.data
+            try:
+                data = self.data
+            except _image_loading_errors:
+                return None
             if data is not None:
                 self._size = data.shape[:2]
         return self._size