Commonsense-Guided VLMs for 2D Fully UnobservedObject Detection

Read full text Summarise with VIRAT
Authors: Ye Liu
Cited by: 0
References: 14

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

This paper focuses on the task of 2D fully unobserved object detection, which requires inferring the presence and spatial distribution of targets that are entirely absent from the visible frame. Despite the potential of Vision-Language Models (VLMs), they often lack the spatial precision required for fine-grained 2D reasoning, while alternative diffusion-based methods incur prohibitive computational costs. To address these challenges, we propose CIUL, a novel reasoning framework that synergizes VLMs with structured prior knowledge through two core innovations: (1) Object-Oriented Commonsense Integration, which constructs an automated knowledge base to provide robust semantic constraints via typical spatial arrangements; and (2) Lightweight In-Context Learning, a paradigm that enables the model to adaptively refine its reasoning logic for unobserved regions using local visual cues from a single image without extensive retraining. Experimental results on the NYU Depth V2 benchmark demonstrate that our approach significantly out-performs existing baselines across key metrics, including 2D Region-wise Accuracy and Normalized Cross-Entropy, effectively bridging the gap between low-level perception and high-level commonsense reasoning.

Fields of Study & Subjects

Multimodal Machine Learning Applications Advanced Neural Network Applications Visual Attention and Saliency Detection

Publication details

Journal / source
Exploring Science Academic Conference Series
Volume
14
Pages
32–40
Publication date
26 March 2026
Language
EN