MAM2020: Eurographics Workshop on Material Appearance Modeling

Permanent URI for this collection

https://diglib.eg.org/handle/10.2312/2632928

Browse

Browsing MAM2020: Eurographics Workshop on Material Appearance Modeling by Subject "I.4.1 [Image Procesing and Computer Vision]"

Now showing 1 - 1 of 1
  • Thumbnail Image
    Item
    The Problem of Entangled Material Properties in SVBRDF Recovery
    (The Eurographics Association, 2020) Saryazdi, Soroush; Murphy, Christian; Mudur, Sudhir; Klein, Reinhard and Rushmeier, Holly
    SVBRDF (spatially varying bidirectional reflectance distribution function) recovery is concerned with deriving the material properties of an object from one or more images. This problem is particularly challenging when the images are casual rather than calibrated captures. It makes the problem highly under specified, since an object can look quite different from different angles and from different light directions. Yet many solutions have been attempted under varying assumptions, and the most promising solutions to date are those which use supervised deep learning techniques. The network is first trained with a large number of synthetically created images of surfaces, usually planar, with known values for material properties and then asked to predict the properties for image(s) of a new object. While the results obtained are impressive as shown through renders of the input object using recovered material properties, there is a problem in the accuracy of the recovered properties. Material properties get entangled, specifically the diffuse and specular reflectance behaviors. Such inaccuracies would hinder various down stream applications which use these properties. In this position paper we present this property entanglement problem. First, we demonstrate the problem through various property map outputs obtained by running a state of the deep learning solution. Next we analyse the present solutions, and argue that the main reason for this entanglement is the way the loss function is defined when training the network. Lastly, we propose potential directions that could be pursued to alleviate this problem.