6‐3: Vision‐Language Models Internalize Human‐Like Memory Colors from Real‐World Objects
SID Symposium Digest of Technical Papers · 2026-08 · 57 (S1) · 49-53
DOI: 10.1002/sdtp.70003_13 ↗Kurz gefasst
Fachartikel in der Zeitschrift SID Symposium Digest of Technical Papers, online erschienen am 11.09.2026. Verfasst von Zhiyu Chen und 2 weiteren Autor:innen. Die Arbeit nennt 25 Literaturangaben.
Abstract (Originalsprache)
Memory colors, the typical perceived colors of familiar objects, profoundly influence human color perception and cognition, yet their systematic representation within modern vision‐language models (VLMs) remains largely unexplored. This study presents the first large‐scale, systematic investigation of memory color representations in VLMs across more than 1,000 everyday objects, extending well beyond traditionally studied color‐diagnostic objects. We introduce two complementary, model‐agnostic probing paradigms: (1) a CLIP‐based semantic similarity framework that quantifies object‐color name associations using 949 XKCD color terms with Gaussian kernel density estimation in CIELAB space, and (2) a direct VLM color‐matching procedure using densely sampled RGB patches that mirrors human psychophysical experiments. Extensive ablation studies over prompt templates, backbone architectures, and color vocabularies demonstrate that VLMs possess rich, human‐like canonical color priors with nearperfect alignment to psychophysical ground truth. Furthermore, we demonstrate two practical applications: a novel CLIP‐based Memory Color Constancy (MCC) algorithm for illuminant estimation, and a VLM‐guided image enhancement pipeline for perceptual optimization. These findings establish that VLMs internalize structured color knowledge closely mirroring human memory colors and can serve as a scalable alternative to laborintensive psychophysical measurements.
Zitieren
APA
Chen, Z., Zhu, C., & Liu, Q. (2026). 6‐3: Vision‐Language Models Internalize Human‐Like Memory Colors from Real‐World Objects. SID Symposium Digest of Technical Papers, 57(S1), 49-53. https://doi.org/10.1002/sdtp.70003_13
Vancouver
Chen Z, Zhu C, Liu Q. 6‐3: Vision‐Language Models Internalize Human‐Like Memory Colors from Real‐World Objects. SID Symposium Digest of Technical Papers. 2026;57(S1):49-53. doi:10.1002/sdtp.70003_13
BibTeX
@article{Chen2026,
author = {Chen, Zhiyu and Zhu, Changrong and Liu, Qiang},
title = {6‐3: Vision‐Language Models Internalize Human‐Like Memory Colors from Real‐World Objects},
journal = {SID Symposium Digest of Technical Papers},
year = {2026},
volume = {57},
number = {S1},
pages = {49--53},
doi = {10.1002/sdtp.70003_13}
}
EUVIMED