[2Q6-GS-10] Vision, speech: Image analysis and application

[2Q6-GS-10-05] Colour-temperature relationship in natural texture images revealed by domain adaptation

〇Hiroki Terashima¹, Hsin-Ni Ho¹, Kouta Wakamatsu², Jinhwan Kwon^3,4, Maki Sakamoto³, Shigeki Nakauchi², Shin'ya Nishida^1,5 (1. NTT Communication Science Laboratories, 2. Toyohashi University of Technology, 3. University of Electro-Communications, 4. Kyoto University of Education, 5. Kyoto University)

Keywords:Temperature, Colour, Domain adaptation

We can visually perceive hotness or coldness of an object without actual touches. The thermal perception of an object is modulated by its colours. How can the colour information predict our visual thermal perception? Previous studies used only a small sample and a single material category. In the present study, we derived a colour-temperature relationship in a data-driven way using 1934 natural texture images of ten material categories. We built models that predict a temperature rating of an image from its colour statistics. To see if the relationship depends on material categories, we extended the feature space using a domain adaptation technique. Training with L1 regularization selected several colour features, mainly a and b in Lab colour space, that are mostly shared by all material categories. The results suggest that the colour-temperature relationship is governed by low-level colour statistics and widely shared by natural material categories.