Understanding color and gloss perception with deep neural networks

Add to your list(s) Download to your calendar using vCal

• Takuma Morimoto, Fellow, Department of Experimental Psychology, University of Oxford, UK
• Thursday 19 September 2024, 14:00-15:00
• FW11 - William Gates Building.

If you have a question about this talk, please contact Rafal Mantiuk.

In daily life, visual properties such as color and gloss are crucial for identifying and distinguishing objects around us. I will present two research projects that investigated the mechanisms of color and gloss perception using deep neural networks (DNNs). In the first project, we used DNNs to explore why we have higher sensitivity to discriminate colors around orange than around purple. We hypothesized that this difference is because our hue discrimination ability is optimized for frequently seen colors in the real world. We approached this by investigating whether similar threshold patterns spontaneously emerge from DNNs trained on over a million real-world color images. We found that human-like hue discrimination ability spontaneously emerges, especially from shallower layers of the network, suggesting that the asymmetry in human color discrimination stems from chromatic regularities in the real world. The second project aimed to identify a computational mechanism underpinning human gloss perception. We conducted online experiments to collect large-scale human gloss judgments for 3,888 object images generated under diverse lighting conditions. We then trained shallow neural networks using these images labeled by human judgments to replicate the judgment patterns. By interpreting the internal architecture that emerged in the network after training, we discovered that a simple computation accounts for human gloss perception. Both projects leverage the power of machine learning to reveal novel underlying mechanisms for the perception of material properties, offering insights that traditional hand-crafted models have not revealed.

Zoom link: https://cam-ac-uk.zoom.us/j/87094914760?pwd=yBmfePQVlg50QNvA9HXvEa9e6Pc4vI.1

This talk is part of the Rainbow Group Seminars series.

This talk is included in these lists:

• All Talks (aka the CURE list)
• Cambridge Centre for Data-Driven Discovery (C2D3)
• Cambridge talks
• Chris Davis' list
• Department of Computer Science and Technology talks and seminars
• FW11 - William Gates Building
• Interested Talks
• Rainbow Group Seminars
• School of Technology
• Trust & Technology Initiative - interesting events
• bld31
• ndk22's list
• ob366-ai4er
• rp587

Note that ex-directory lists are not shown.