The Semantic PASCAL-Part Dataset

<p><strong>The Semantic PASCAL-Part dataset</strong></p> <p>The Semantic PASCAL-Part dataset is the RDF version of the famous PASCAL-Part dataset used for object detection in Computer Vision. Each image is annotated with <strong>bounding boxes</strong> containing a single object. Couples of bounding boxes are annotated with the part-whole relationship. For example, the bounding box of a car has the part-whole annotation with the bounding boxes of its wheels.</p> <p>This original release joins Computer Vision with Semantic Web as the objects in the dataset are aligned with concepts from:</p> <ul> <li>the provided supporting ontology;</li> <li>the <a href="https://wordnet.princeton.edu/">WordNet</a> database through its synstes;</li> <li>the <a href="https://yago-knowledge.org/">Yago</a> ontology.</li> </ul> <p>The provided Python 3 code (see the <a href="https://github.com/ivanDonadello/semantic-PASCAL-Part">GitHub repo</a>) is able to browse the dataset and convert it in RDF knowledge graph format. This new format easily allows the fostering of research in both Semantic Web and Machine Learning fields.</p> <p><strong>Structure of the semantic PASCAL-Part Dataset</strong></p> <p>This is the folder structure of the dataset:</p> <ul> <li><code>semanticPascalPart</code>: it contains the refined images and annotations (e.g., small specific parts are merged into bigger parts) of the PASCAL-Part dataset in Pascal-voc style. <ul> <li><code>Annotations_set</code>: the test set annotations in <code>.xml</code> format. For further information See the PASCAL VOC format <a href="http://host.robots.ox.ac.uk/pascal/VOC/index.html">here</a>.</li> <li><code>Annotations_trainval</code>: the train and validation set annotations in <code>.xml</code> format. For further information See the PASCAL VOC format <a href="http://host.robots.ox.ac.uk/pascal/VOC/index.html">here</a>.</li> <li><code>JPEGImages_test</code>: the test set images in <code>.jpg</code> format.</li> <li><code>JPEGImages_trainval</code>: the train and validation set images in <code>.jpg</code> format.</li> <li><code>test.txt</code>: the 2416 image filenames in the test set.</li> <li><code>trainval.txt</code>: the 7687 image filenames in the train and validation set.</li> </ul> </li> </ul> <p><strong>The PASCAL-Part Ontology</strong></p> <p>The PASCAL-Part OWL ontology formalizes, through logical axioms, the part-of relationship between whole objects (22 classes) and their parts (39 classes). The ontology contains 85 logical axiomns in Description Logic in (for example) the following form:</p> <pre><code>Every potted_plant has exactly 1 plant AND has exactly 1 pot </code></pre> <p>We provide two versions of the ontology: with and without cardinality constraints in order to allow users to experiment with or without them. The WordNet alignment is encoded in the ontology as annotations. We further provide the <code>WordNet_Yago_alignment.csv</code> file with both WordNet and Yago alignments.</p> <p>The ontology can be browsed with many Semantic Web tools such as:</p> <ul> <li><a href="https://protege.stanford.edu/">Prot&eacute;g&eacute;</a>: a graphical tool for ongology modelling;</li> <li><a href="http://owlapi.sourceforge.net/">OWLAPI</a>: Java API for manipulating OWL ontologies;</li> <li><a href="https://rdflib.readthedocs.io/en/stable/">rdflib</a>: Python API for working with the RDF format.</li> <li>RDF stores: databases for storing and semantically retrieve RDF triples. See <a href="https://www.w3.org/wiki/LargeTripleStores">here</a> for some examples.</li> </ul> <p><strong>Citing semantic PASCAL-Part</strong></p> <p>If you use semantic PASCAL-Part in your research, please use the following BibTeX entry</p> <pre><code>@article{DBLP:journals/ia/DonadelloS16, author = {Ivan Donadello and Luciano Serafini}, title = {Integration of numeric and symbolic information for semantic image interpretation}, journal = {Intelligenza Artificiale}, volume = {10}, number = {1}, pages = {33--47}, year = {2016} } </code></pre>