Inproceedings
Semantic View Induction for Object-Guided Scene Structuring
Opeyemi Bamigbade; Mark Scanlon; John Sheppard. Proceedings of the Computer Vision Conference (CVC) 2027, 2027.
Contribution Summary
This paper examines whether reorganising scene datasets around recurring visual configurations can improve downstream learning without changing category labels. Semantic View Induction discovers latent views from image representations without predefined subclasses or manual view annotations. SceneStruct then uses open-set object evidence to assign images to those views while retaining samples whose evidence is ambiguous. Experiments identify three views in Hotel-ID and two in SUN397, confidently structuring 94.08% and 99.38% of samples respectively. On Hotel-ID, the reported retrieval P@1 rises from 38.60% to 48.28%, mAP@5 from 49.80% to 57.17%, and classification accuracy from 36.56% to 40.96%. SUN397 classification accuracy rises from 63.41% to 69.74%. The contribution is a data-centric framework for exposing latent scene structure, with explicit handling of uncertainty and empirical evaluation of retrieval and classification outcomes.
Keywords: Scene Understanding; Data-Centric Learning; Open-Vocabulary Object Detection; Semantic Induction
Abstract
Scene datasets often exhibit substantial intra-class variability, where images belonging to the same semantic category contain multiple latent scene configurations characterised by different object compositions, spatial layouts, and functional contexts. This variability degrades the underlying semantic structure and increases representation ambiguity. This work introduces a two-stage scene structuring framework composed of Semantic View Induction (SVI) and SceneStruct. SVI discovers latent semantic views directly from image representations without predefined subclasses or manual annotations. SceneStruct then assigns images to the induced views using open-set object evidence while explicitly preserving ambiguous samples. Experimental evaluation on Hotel-ID and SUN397 demonstrates that the proposed framework induces three and two semantic views, respectively, while confidently structuring 94.08% and 99.38% of samples, leaving 5.92% and 0.61% as ambiguous. On Hotel-ID, restructuring increases retrieval P@1 from 38.60% to 48.28% and mAP@5 from 49.80% to 57.17%, while classification accuracy increases from 36.56% to 40.96%. On SUN397, classification accuracy increases from 63.41% to 69.74%. These results show that semantic view induction can uncover latent scene structure while preserving uncertain samples and is associated with improved downstream performance.
BibTeX
@InProceedings{bamigbade2027Object-GuidedSceneStructuring,
author="Opeyemi Bamigbade and Mark Scanlon and John Sheppard",
title="Semantic View Induction for Object-Guided Scene Structuring",
booktitle="Proceedings of the Computer Vision Conference (CVC) 2027",
year="2027",
publisher="Springer Nature Switzerland",
address="Cham",
pages="",
abstract="Scene datasets often exhibit substantial intra-class variability, where images belonging to the same semantic category contain multiple latent scene configurations characterised by different object compositions, spatial layouts, and functional contexts. This variability degrades the underlying semantic structure and increases representation ambiguity. This work introduces a two-stage scene structuring framework composed of Semantic View Induction (SVI) and SceneStruct. SVI discovers latent semantic views directly from image representations without predefined subclasses or manual annotations. SceneStruct then assigns images to the induced views using open-set object evidence while explicitly preserving ambiguous samples. Experimental evaluation on Hotel-ID and SUN397 demonstrates that the proposed framework induces three and two semantic views, respectively, while confidently structuring 94.08% and 99.38% of samples, leaving 5.92% and 0.61% as ambiguous. On Hotel-ID, restructuring increases retrieval P@1 from 38.60% to 48.28% and mAP@5 from 49.80% to 57.17%, while classification accuracy increases from 36.56% to 40.96%. On SUN397, classification accuracy increases from 63.41% to 69.74%. These results show that semantic view induction can uncover latent scene structure while preserving uncertain samples and is associated with improved downstream performance."
}