Merge pull request #5349 from enricoschroeder:improve-oxford_iiit_pet

The TensorFlow Datasets Authors · The TensorFlow Datasets Authors · commit 16559144218e · 2024-08-08T05:56:33.000-07:00
PiperOrigin-RevId: 660802201
diff --git a/tensorflow_datasets/datasets/oxford_iiit_pet/README.md b/tensorflow_datasets/datasets/oxford_iiit_pet/README.md
@@ -1,3 +1,6 @@
 The Oxford-IIIT pet dataset is a 37 category pet image dataset with roughly 200
 images for each class. The images have large variations in scale, pose and
-lighting. All images have an associated ground truth annotation of breed.
+lighting. All images have an associated ground truth annotation of breed and
+species. Additionally, head bounding boxes are provided for the training split,
+allowing using this dataset for simple object detection tasks. In the test
+split, the bounding boxes are empty.
diff --git a/tensorflow_datasets/datasets/oxford_iiit_pet/checksums.tsv b/tensorflow_datasets/datasets/oxford_iiit_pet/checksums.tsv
@@ -1,2 +1,2 @@
-http://www.robots.ox.ac.uk/~vgg/data/pets/data/annotations.tar.gz	19173078	52425fb6de5c424942b7626b428656fcbd798db970a937df61750c0f1d358e91	annotations.tar.gz
-http://www.robots.ox.ac.uk/~vgg/data/pets/data/images.tar.gz	791918971	67195c5e1c01f1ab5f9b6a5d22b8c27a580d896ece458917e61d459337fa318d	images.tar.gz
+https://thor.robots.ox.ac.uk/~vgg/data/pets/annotations.tar.gz	19173078	52425fb6de5c424942b7626b428656fcbd798db970a937df61750c0f1d358e91	annotations.tar.gz
+https://thor.robots.ox.ac.uk/~vgg/data/pets/images.tar.gz	791918971	67195c5e1c01f1ab5f9b6a5d22b8c27a580d896ece458917e61d459337fa318d	images.tar.gz
diff --git a/tensorflow_datasets/datasets/oxford_iiit_pet/dummy_data/annotations/xmls/image0_0.xml b/tensorflow_datasets/datasets/oxford_iiit_pet/dummy_data/annotations/xmls/image0_0.xml
@@ -0,0 +1 @@
+<annotation><folder>OXIIIT</folder><filename>Abyssinian_1.jpg</filename><source><database>OXFORD-IIIT Pet Dataset</database><annotation>OXIIIT</annotation><image>flickr</image></source><size><width>600</width><height>400</height><depth>3</depth></size><segmented>0</segmented><object><name>cat</name><pose>Frontal</pose><truncated>0</truncated><occluded>0</occluded><bndbox><xmin>333</xmin><ymin>72</ymin><xmax>425</xmax><ymax>158</ymax></bndbox><difficult>0</difficult></object></annotation>
diff --git a/tensorflow_datasets/datasets/oxford_iiit_pet/dummy_data/annotations/xmls/image0_1.xml b/tensorflow_datasets/datasets/oxford_iiit_pet/dummy_data/annotations/xmls/image0_1.xml
@@ -0,0 +1 @@
+<annotation><folder>OXIIIT</folder><filename>Abyssinian_10.jpg</filename><source><database>OXFORD-IIIT Pet Dataset</database><annotation>OXIIIT</annotation><image>flickr</image></source><size><width>375</width><height>500</height><depth>3</depth></size><segmented>0</segmented><object><name>cat</name><pose>Frontal</pose><truncated>0</truncated><occluded>0</occluded><bndbox><xmin>72</xmin><ymin>105</ymin><xmax>288</xmax><ymax>291</ymax></bndbox><difficult>0</difficult></object></annotation>
diff --git a/tensorflow_datasets/datasets/oxford_iiit_pet/dummy_data/annotations/xmls/image1_0.xml b/tensorflow_datasets/datasets/oxford_iiit_pet/dummy_data/annotations/xmls/image1_0.xml
@@ -0,0 +1 @@
+<annotation><folder>OXIIIT</folder><filename>Abyssinian_11.jpg</filename><source><database>OXFORD-IIIT Pet Dataset</database><annotation>OXIIIT</annotation><image>flickr</image></source><size><width>192</width><height>288</height><depth>3</depth></size><segmented>0</segmented><object><name>cat</name><pose>Frontal</pose><truncated>0</truncated><occluded>0</occluded><bndbox><xmin>43</xmin><ymin>57</ymin><xmax>94</xmax><ymax>105</ymax></bndbox><difficult>0</difficult></object></annotation>
diff --git a/tensorflow_datasets/datasets/oxford_iiit_pet/dummy_data/annotations/xmls/image1_1.xml b/tensorflow_datasets/datasets/oxford_iiit_pet/dummy_data/annotations/xmls/image1_1.xml
@@ -0,0 +1 @@
+<annotation><folder>OXIIIT</folder><filename>Abyssinian_18.jpg</filename><source><database>OXFORD-IIIT Pet Dataset</database><annotation>OXIIIT</annotation><image>flickr</image></source><size><width>335</width><height>500</height><depth>3</depth></size><segmented>0</segmented><object><name>cat</name><pose>Frontal</pose><truncated>0</truncated><occluded>0</occluded><bndbox><xmin>135</xmin><ymin>87</ymin><xmax>314</xmax><ymax>285</ymax></bndbox><difficult>0</difficult></object></annotation>
diff --git a/tensorflow_datasets/datasets/oxford_iiit_pet/dummy_data/annotations/xmls/image2_0.xml b/tensorflow_datasets/datasets/oxford_iiit_pet/dummy_data/annotations/xmls/image2_0.xml
@@ -0,0 +1 @@
+<annotation><folder>OXIIIT</folder><filename>Abyssinian_17.jpg</filename><source><database>OXFORD-IIIT Pet Dataset</database><annotation>OXIIIT</annotation><image>flickr</image></source><size><width>275</width><height>411</height><depth>3</depth></size><segmented>0</segmented><object><name>cat</name><pose>Frontal</pose><truncated>0</truncated><occluded>0</occluded><bndbox><xmin>38</xmin><ymin>92</ymin><xmax>136</xmax><ymax>192</ymax></bndbox><difficult>0</difficult></object></annotation>
diff --git a/tensorflow_datasets/datasets/oxford_iiit_pet/dummy_data/annotations/xmls/image2_1.xml b/tensorflow_datasets/datasets/oxford_iiit_pet/dummy_data/annotations/xmls/image2_1.xml
@@ -0,0 +1 @@
+<annotation><folder>OXIIIT</folder><filename>Abyssinian_16.jpg</filename><source><database>OXFORD-IIIT Pet Dataset</database><annotation>OXIIIT</annotation><image>flickr</image></source><size><width>500</width><height>335</height><depth>3</depth></size><segmented>0</segmented><object><name>cat</name><pose>Frontal</pose><truncated>0</truncated><occluded>0</occluded><bndbox><xmin>375</xmin><ymin>53</ymin><xmax>458</xmax><ymax>134</ymax></bndbox><difficult>0</difficult></object></annotation>
diff --git a/tensorflow_datasets/datasets/oxford_iiit_pet/dummy_data/annotations/xmls/image3_0.xml b/tensorflow_datasets/datasets/oxford_iiit_pet/dummy_data/annotations/xmls/image3_0.xml
@@ -0,0 +1 @@
+<annotation><folder>OXIIIT</folder><filename>Abyssinian_14.jpg</filename><source><database>OXFORD-IIIT Pet Dataset</database><annotation>OXIIIT</annotation><image>flickr</image></source><size><width>335</width><height>500</height><depth>3</depth></size><segmented>0</segmented><object><name>cat</name><pose>Frontal</pose><truncated>0</truncated><occluded>0</occluded><bndbox><xmin>164</xmin><ymin>71</ymin><xmax>252</xmax><ymax>163</ymax></bndbox><difficult>0</difficult></object></annotation>
diff --git a/tensorflow_datasets/datasets/oxford_iiit_pet/dummy_data/annotations/xmls/image3_1.xml b/tensorflow_datasets/datasets/oxford_iiit_pet/dummy_data/annotations/xmls/image3_1.xml
@@ -0,0 +1 @@
+<annotation><folder>OXIIIT</folder><filename>Abyssinian_15.jpg</filename><source><database>OXFORD-IIIT Pet Dataset</database><annotation>OXIIIT</annotation><image>flickr</image></source><size><width>477</width><height>422</height><depth>3</depth></size><segmented>0</segmented><object><name>cat</name><pose>Frontal</pose><truncated>0</truncated><occluded>0</occluded><bndbox><xmin>16</xmin><ymin>16</ymin><xmax>431</xmax><ymax>357</ymax></bndbox><difficult>0</difficult></object></annotation>
diff --git a/tensorflow_datasets/datasets/oxford_iiit_pet/dummy_data/annotations/xmls/image4_0.xml b/tensorflow_datasets/datasets/oxford_iiit_pet/dummy_data/annotations/xmls/image4_0.xml
@@ -0,0 +1 @@
+<annotation><folder>OXIIIT</folder><filename>Abyssinian_13.jpg</filename><source><database>OXFORD-IIIT Pet Dataset</database><annotation>OXIIIT</annotation><image>flickr</image></source><size><width>245</width><height>280</height><depth>3</depth></size><segmented>0</segmented><object><name>cat</name><pose>Frontal</pose><truncated>0</truncated><occluded>0</occluded><bndbox><xmin>145</xmin><ymin>25</ymin><xmax>218</xmax><ymax>96</ymax></bndbox><difficult>0</difficult></object></annotation>
diff --git a/tensorflow_datasets/datasets/oxford_iiit_pet/dummy_data/annotations/xmls/image4_1.xml b/tensorflow_datasets/datasets/oxford_iiit_pet/dummy_data/annotations/xmls/image4_1.xml
@@ -0,0 +1 @@
+<annotation><folder>OXIIIT</folder><filename>Abyssinian_12.jpg</filename><source><database>OXFORD-IIIT Pet Dataset</database><annotation>OXIIIT</annotation><image>flickr</image></source><size><width>335</width><height>500</height><depth>3</depth></size><segmented>0</segmented><object><name>cat</name><pose>Frontal</pose><truncated>0</truncated><occluded>0</occluded><bndbox><xmin>94</xmin><ymin>83</ymin><xmax>211</xmax><ymax>190</ymax></bndbox><difficult>0</difficult></object></annotation>
diff --git a/tensorflow_datasets/datasets/oxford_iiit_pet/oxford_iiit_pet_dataset_builder.py b/tensorflow_datasets/datasets/oxford_iiit_pet/oxford_iiit_pet_dataset_builder.py
@@ -16,11 +16,13 @@
 """Oxford-IIIT pet dataset."""
 
 import os
+import xml.etree.ElementTree as ET
 
+from etils import epath
 from tensorflow_datasets.core.utils.lazy_imports_utils import tensorflow as tf
 import tensorflow_datasets.public_api as tfds
 
-_BASE_URL = "http://www.robots.ox.ac.uk/~vgg/data/pets/data"
+_BASE_URL = "https://thor.robots.ox.ac.uk/~vgg/data/pets"
 
 _LABEL_CLASSES = [
     "Abyssinian",
@@ -63,11 +65,56 @@
 ]
 _SPECIES_CLASSES = ["Cat", "Dog"]
 
+# List of samples with corrupt image files (mostly wrong format -> we are fixing
+# these during dataset creation)
+_CORRUPT_SAMPLES = [
+    "beagle_116",
+    "chihuahua_121",
+    "Abyssinian_5",
+    "Abyssinian_34",
+    "Egyptian_Mau_14",
+    "Egyptian_Mau_139",
+    "Egyptian_Mau_145",
+    "Egyptian_Mau_156",
+    "Egyptian_Mau_167",
+    "Egyptian_Mau_177",
+    "Egyptian_Mau_186",
+    "Egyptian_Mau_191",
+]
+
+_EMPTY_BBOX = tfds.features.BBox(0.0, 0.0, 0.0, 0.0)
+
+
+def _get_head_bbox(annotations_filepath):
+  """Read head bbox from annotation XML file."""
+  with open(annotations_filepath) as f:
+    root = ET.parse(f).getroot()
+
+    size = root.find("size")
+    width = float(size.find("width").text)  # pytype: disable=attribute-error
+    height = float(size.find("height").text)  # pytype: disable=attribute-error
+
+    obj = root.find("object")
+    bndbox = obj.find("bndbox")  # pytype: disable=attribute-error
+    xmax = float(bndbox.find("xmax").text)  # pytype: disable=attribute-error
+    xmin = float(bndbox.find("xmin").text)  # pytype: disable=attribute-error
+    ymax = float(bndbox.find("ymax").text)  # pytype: disable=attribute-error
+    ymin = float(bndbox.find("ymin").text)  # pytype: disable=attribute-error
+
+    return tfds.features.BBox(
+        ymin / height, xmin / width, ymax / height, xmax / width
+    )
+
 
 class Builder(tfds.core.GeneratorBasedBuilder):
   """Oxford-IIIT pet dataset."""
 
-  VERSION = tfds.core.Version("3.2.0")
+  VERSION = tfds.core.Version("4.0.0")
+  RELEASE_NOTES = {
+      "4.0.0": (
+          "Add head bounding boxes. Fix corrupt images. Update dataset URL."
+      )
+  }
 
   def _info(self):
     return self.dataset_info_from_configs(
@@ -79,6 +126,7 @@ def _info(self):
             "segmentation_mask": tfds.features.Image(
                 shape=(None, None, 1), use_colormap=True
             ),
+            "head_bbox": tfds.features.BBoxFeature(),
         }),
         supervised_keys=("image", "label"),
         homepage="http://www.robots.ox.ac.uk/~vgg/data/pets/",
@@ -122,22 +170,48 @@ def _split_generators(self, dl_manager):
   def _generate_examples(
       self, images_dir_path, annotations_dir_path, images_list_file
   ):
-    with tf.io.gfile.GFile(images_list_file, "r") as images_list:
+    with epath.Path(images_list_file).open("r") as images_list:
       for line in images_list:
         image_name, label, species, _ = line.strip().split(" ")
 
+        image_path = os.path.join(images_dir_path, image_name + ".jpg")
+
+        if image_name in _CORRUPT_SAMPLES:
+          # some images caused 'Corrupt JPEG data...' messages during training
+          # or any other iteration recoding them once fixes the issue
+          # (discussion: https://github.com/tensorflow/datasets/issues/2188)
+          with epath.Path(image_path).open("rb") as image_file:
+            img_data = image_file.read()
+            img_tensor = tf.image.decode_image(img_data)
+            if (
+                tf.shape(img_tensor)[-1] == 4
+            ):  # some files have an alpha channel -> remove
+              img_tensor = img_tensor[:, :, :-1]
+            img_recoded = tf.io.encode_jpeg(img_tensor)
+          with epath.Path(image_path).open("wb") as image_file:
+            image_file.write(img_recoded.numpy())
+
         trimaps_dir_path = os.path.join(annotations_dir_path, "trimaps")
+        xmls_dir_path = os.path.join(annotations_dir_path, "xmls")
 
         trimap_name = image_name + ".png"
+        xml_name = image_name + ".xml"
         image_name += ".jpg"
         label = int(label) - 1
         species = int(species) - 1
 
+        try:
+          head_bbox = _get_head_bbox(os.path.join(xmls_dir_path, xml_name))
+        except FileNotFoundError:
+          # test samples do not have an annotation file
+          head_bbox = _EMPTY_BBOX
+
         record = {
             "image": os.path.join(images_dir_path, image_name),
             "label": int(label),
             "species": species,
             "file_name": image_name,
             "segmentation_mask": os.path.join(trimaps_dir_path, trimap_name),
+            "head_bbox": head_bbox,
         }
         yield image_name, record

Original file line number	Diff line number	Diff line change
`@@ -0,0 +1 @@`
	`1`	`+<annotation><folder>OXIIIT</folder><filename>Abyssinian_1.jpg</filename><source><database>OXFORD-IIIT Pet Dataset</database><annotation>OXIIIT</annotation><image>flickr</image></source><size><width>600</width><height>400</height><depth>3</depth></size><segmented>0</segmented><object><name>cat</name><pose>Frontal</pose><truncated>0</truncated><occluded>0</occluded><bndbox><xmin>333</xmin><ymin>72</ymin><xmax>425</xmax><ymax>158</ymax></bndbox><difficult>0</difficult></object></annotation>`