Suppr超能文献

“哪里”是什么:物理推理为物体定位提供依据。

What is "Where": Physical Reasoning Informs Object Location.

作者信息

Boger Tal, Ullman Tomer

机构信息

Department of Psychology, Yale University, New Haven, CT, USA.

Department of Psychology, Harvard University, Cambridge, MA, USA.

出版信息

Open Mind (Camb). 2023 May 1;7:130-140. doi: 10.1162/opmi_a_00075. eCollection 2023.

Abstract

A central puzzle the visual system tries to solve is: "what is where?" While a great deal of research attempts to model object recognition ("what"), a comparatively smaller body of work seeks to model object location ("where"), especially in perceiving everyday objects. How do people locate an object, right now, in front of them? In three experiments collecting over 35,000 judgements on stimuli spanning different levels of realism (line drawings, real images, and crude forms), participants clicked "where" an object is, as if pointing to it. We modeled their responses with eight different methods, including both human response-based models (judgements of physical reasoning, spatial memory, free-response "click anywhere" judgements, and judgements of where people would grab the object), and image-based models (uniform distributions over the image, convex hull, saliency map, and medial axis). Physical reasoning was the best predictor of "where," performing significantly better than even spatial memory and free-response judgements. Our results offer insight into the perception of object locations while also raising interesting questions about the relationship between physical reasoning and visual perception.

摘要

视觉系统试图解决的一个核心难题是

“什么东西在哪里?”虽然大量研究致力于对物体识别(“什么”)进行建模,但相对较少的研究工作旨在对物体位置(“哪里”)进行建模,尤其是在感知日常物体方面。人们如何在此时此刻找到眼前的一个物体呢?在三项实验中,针对跨越不同逼真程度(线条图、真实图像和粗略图形)的刺激收集了超过35000份判断,参与者点击物体所在的“位置”,就好像在指向它一样。我们用八种不同方法对他们的反应进行建模,包括基于人类反应的模型(物理推理判断、空间记忆、自由反应“任意点击”判断以及人们抓取物体位置的判断)和基于图像的模型(图像上的均匀分布、凸包、显著性图和中轴线)。物理推理是“哪里”的最佳预测指标,其表现甚至显著优于空间记忆和自由反应判断。我们的研究结果为物体位置的感知提供了见解,同时也引发了关于物理推理与视觉感知之间关系的有趣问题。

https://cdn.ncbi.nlm.nih.gov/pmc/blobs/3664/10320814/97621b0d248a/opmi-07-130-g001.jpg

文献AI研究员

20分钟写一篇综述,助力文献阅读效率提升50倍。

立即体验

用中文搜PubMed

大模型驱动的PubMed中文搜索引擎

马上搜索

文档翻译

学术文献翻译模型,支持多种主流文档格式。

立即体验