Publication

Abstract : High-level image understanding includes phases like object detection, predicate classification, and attribute classification. The outputs from each phase are merged to build a scene graph, which arranges the elements in a structured manner. Scene graphs have shown their proficiency in various tasks like image retrieval, visual question answering, and image generation. However, data is an essential aspect for such tasks, especially when the models are too complex. We introduce Compact-VG, a refined subset of the popular dataset visual genome. This subset contains 200 object categories, 10 predicates, and 16 attributes. Studies show that, even when we consider only the most common categories of objects, predicates, and attributes, the extracted dataset is still very rich, with a mean of 14.1 objects, 18.5 attributes, and 19.7 relationships per image.