Improve oxford_iiit_pet dataset:

enricoschroeder · enricoschroeder · commit 1ff3092ebbb9 · 2024-04-04T16:10:19.000+02:00
- Include head bounding boxes for the training split.
- Exclude samples with invalid or corrupt image files (preventing 'Corrupt JPEG data:...' error messages while using the dataset).
diff --git a/tensorflow_datasets/datasets/oxford_iiit_pet/README.md b/tensorflow_datasets/datasets/oxford_iiit_pet/README.md
@@ -1,3 +1,7 @@
 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.
+
+Please take note of the following:
+- Valid bounding boxes of the pets' heads are only provided for the training split. The test split does not contain annotations.
+- A few image files in the original dataset are corrupt or in a wrong format. These are being skipped. 
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 +16,12 @@
 """Oxford-IIIT pet dataset."""
 
 import os
+import xml.etree.ElementTree
 
 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",
@@ -63,6 +64,44 @@
 ]
 _SPECIES_CLASSES = ["Cat", "Dog"]
 
+# List of samples with corrupt image files
+_SKIP_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"
+]
+
+def _get_head_bbox(annon_filepath):
+  """Read head bbox from annotation XML file."""
+  with tf.io.gfile.GFile(annon_filepath, "r") as f:
+    root = xml.etree.ElementTree.parse(f).getroot()
+
+    # Disable pytype to avoid attribute-error due to find returning
+    # Optional[Element]
+    # pytype: disable=attribute-error
+    size = root.find("size")
+    width = float(size.find("width").text)
+    height = float(size.find("height").text)
+
+    obj = root.find("object")
+    bndbox = obj.find("bndbox")
+    xmax = float(bndbox.find("xmax").text)
+    xmin = float(bndbox.find("xmin").text)
+    ymax = float(bndbox.find("ymax").text)
+    ymin = float(bndbox.find("ymin").text)
+    return tfds.features.BBox(
+              ymin / height, xmin / width, ymax / height, xmax / width
+        )
+
 
 class Builder(tfds.core.GeneratorBasedBuilder):
   """Oxford-IIIT pet dataset."""
@@ -79,6 +118,7 @@ def _info(self):
             "segmentation_mask": tfds.features.Image(
                 shape=(None, None, 1), use_colormap=True
             ),
+            "head": tfds.features.BBoxFeature()
         }),
         supervised_keys=("image", "label"),
         homepage="http://www.robots.ox.ac.uk/~vgg/data/pets/",
@@ -126,18 +166,31 @@ def _generate_examples(
       for line in images_list:
         image_name, label, species, _ = line.strip().split(" ")
 
+        # skip corrupt samples
+        if image_name in _SKIP_SAMPLES:
+          continue
+
         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 tf.errors.NotFoundError:
+            # test samples do not have an annotation file
+            head_bbox = tfds.features.BBox(0., 0., 0., 0.)
+
         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": 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>`