December 2022
Volume 22, Issue 14
Open Access
Vision Sciences Society Annual Meeting Abstract  |   December 2022
Replaceability of two deep generative models trained with a pair of translucent objects with different geometries
Author Affiliations & Notes
  • Masataka Sawayama
    Inria
  • Chenxi Liao
    American University
  • Shin'ya Nishida
    Cognitive Informatics Lab, Dept. of Intelligence Science and Technology, Graduate School of Informatics, Kyoto University
    NTT Communication Science Labs, Nippon Telegraph and Telephone Corp.
  • Bei Xiao
    American University
  • Footnotes
    Acknowledgements  This work was supported by JSPS Kakenhi Grant Number JP20H05957
Journal of Vision December 2022, Vol.22, 3522. doi:https://doi.org/10.1167/jov.22.14.3522
  • Views
  • Share
  • Tools
    • Alerts
      ×
      This feature is available to authenticated users only.
      Sign In or Create an Account ×
    • Get Citation

      Masataka Sawayama, Chenxi Liao, Shin'ya Nishida, Bei Xiao; Replaceability of two deep generative models trained with a pair of translucent objects with different geometries. Journal of Vision 2022;22(14):3522. https://doi.org/10.1167/jov.22.14.3522.

      Download citation file:


      © ARVO (1962-2015); The Authors (2016-present)

      ×
  • Supplements
Abstract

Perceived translucency of objects with various geometries can be different, even when physical translucency is kept constant. The logic of linking translucent images with different geometries, e.g., sharp vs. rounded corner edges, has been missing. For instance, can the translucent appearance of an object with a sharp corner edge be synthesized from another object without it? One possibility is that different geometries share common latent image factors diagnostic for perceived translucency, and each geometry affects the latent image factors in a non-categorical manner. We investigate this possibility by exploring the functional replaceability of paired deep generative models (e.g., VAEs or GANs) trained with rendered translucent images of paired geometries. If the paired geometries have a continuous effect on translucent images, the latent codes of one network should be replaceable, i.e., linearly transferable to the other network. Specifically, in one condition, one side of a paired network was trained with the translucent images of a "cube" geometry rendered with many combinations of physical parameters such as extinction coefficients, albedos, lightings, and phase functions. The other side was separately trained using a "bunny" geometry with the same physical parameters. Subsequently, we trained linear regression decoders to transform the latent codes of one network into the other. We tested whether the input of one geometry (e.g., "cube") can produce the output of the other geometry (e.g., "bunny") through the regression of the latent codes. Our results using ten natural geometries (Che et al., 2020, IEEE ICCP) showed that all geometry pairs could be replaceable in terms of various image evaluation metrics (PSNR, LPIPS, and DISTS). Furthermore, changing the scale parameter of the regression could quantitatively alter the perceived translucency of the output image. These findings suggest that the ten natural geometries have common image features for translucent images, which are continuously controllable.

×
×

This PDF is available to Subscribers Only

Sign in or purchase a subscription to access this content. ×

You must be signed into an individual account to use this feature.

×