projects people publications datasets

Learning Models for Illumination, Shadows and Shading

Pictures result from the complex interactions between the geometry, illumination and materials present in a scene. Estimating any of these factors individually from a single image leads to severely underconstrained problems.

Having a rough geometric model of the scene has allowed us to jointly estimate the illumination and the cast shadows of the scene. We solve the joint estimation using graphical models and novel shadow cues such as the bright channel. We further develop our method so as to be able to perform inference on not only illumination and cast shadows but also on the scene's geometry, thus refining the initial rough geometrical model.

Scene understanding methods that want to account for illumination effects require reasonable shadow estimates. General shadow detection from a single image is a challenging problem, especially when dealing with consumer grade pictures (or pictures from the internet). We investigate methods for single image shadow detection based on learning the appearance of shadows from labelled data sets. Our methods are the state of the art in shadow detection.

In our work we take advantage of domain knowledge in illumination and image formation for feature design, as well as extensive use of machine learning techniques such as graphical models (MRFs, CRFs) and classifiers (SVMs).

Ongoing lines of work are: illumination based scene understanding, shadow removal and shadow segmentation.

  • Students:
    • Tomas F. Yago Vicente
  • Former students:
    • Alexandros Panagopoulos
  • The Photometry of Intrinsic Images


    Intrinsic characterization of scenes is often the best way to overcome the illumination variability artifacts that com- plicate most computer vision problems, from 3D reconstruction to object or material recognition. This paper examines the deficiency of existing intrinsic image models to accurately account for the effects of illuminant color and sensor characteristics in the estimation of intrinsic images and presents a generic framework which incorporates insights from color constancy research to the intrinsic image decomposition problem. The proposed mathematical formulation includes information about the color of the illuminant and the effects of the camera sensors, both of which modify the observed color of the reflectance of the objects in the scene during the acquisition process. By modeling these effects, we get a “truly intrinsic” reflectance image, which we call absolute reflectance, which is invariant to changes of illuminant or camera sensors. This model allows us to represent a wide range of intrinsic image decompositions depending on the specific assumptions on the geometric properties of the scene configuration and the spectral properties of the light source and the acquisition system, thus unifying previous models in a single general framework. We demonstrate that even partial information about sensors improves significantly the estimated reflectance images, thus making our method applicable for a wide range of sensors. We validate our general intrinsic image framework experimentally with both synthetic data and natural images.


    Simultaneous Cast Shadows, Illumination and Geometry Inference Using Hypergraphs


    The cast shadows in an image provide important information about illumination and geometry. In this paper, we utilize this information in a novel framework in order to jointly recover the illumination environment, a set of geometry parameters, and an estimate of the cast shadows in the scene given a single image and coarse initial 3D geometry. We model the interaction of illumination and geometry in the scene and associate it with image evidence for cast shadows using a higher order Markov Random Field (MRF) illumination model, while we also introduce a method to obtain approximate image evidence for cast shadows. Capturing the interaction between light sources and geometry in the proposed graphical model necessitates higher order cliques and continuous- valued variables, which make inference challenging. Taking advantage of domain knowledge, we provide a two-stage minimization technique for the MRF energy of our model. We evaluate our method in different datasets, both synthetic and real. Our model is robust to rough knowledge of geometry and inaccurate initial shadow estimates, allowing a generic coarse 3D model to represent a whole class of objects for the task of illumination estimation, or the estimation of geometry parameters to refine our initial knowledge of scene geometry, simultaneously with illumination estimation.


    Illumination Modeling for Computer Vision & Computer Graphics

    We are researching problems related with the interaction of illumination and 3D shape in images for Computer Vision (shape estimation, tracking, recognition) and Computer Graphics (image relighting, augmented reality), in which object appearance is significantly influenced by the illumination conditions of the scene. Optical images are formed by measurements of the light that reflects from surfaces of various materials and orientations in 3D space. Effective approaches to numerous problems related to visual appearance, from photo-realistic image synthesis, to convincing Augmented Reality environments to object recognition, rely on the accuracy and efficiency of models of visual appearance of objects, thus motivating the study of the image formation process. In addition to knowing the geometry of the scene and of the cameras, we also need to model the reflectance properties of the surfaces being imaged and the configuration of illumination sources in the scene. In the general case it is impossible to determine simultaneously shape, reflectance and illumination without having any prior knowledge of the scene, so we have been working on a number of subproblems where different priors are used.

    We have also developed novel solutions for 3D Shape estimation from shading from single images and href="content/papers/2000/samaras-cvpr2000.pdf">stereo sets, multiple light source estimation, visual tracking of deformable objects with little texture, and face recognition under arbitrary illumination.

  • Former students:
    • Yang Wang

  • In Dr. Samaras's Ph.D. thesis, he proposed a method for the incorporation of any type of illumination constraints in deformable model frameworks. If the light source direction is unknown, both light source and object shape can be recovered, in . The applicability of the above methods was extended with the integration of shape from shading (SFS) and stereo, for non-constant albedo and non-uniformly Lambertian surfaces. In the case of moving objects, optical flow becomes part of a generalized illumination constraint, which was used for visual tracking of deformable objects with little texture.

    Often it is impossible to know exactly either the reflectance properties or the shape of an object, but we can have a statistical prior for them, especially nowadays that image databases are ubiquitous. This led us to study the statistical properties of shape and illumination for the case of particular object categories. When applied to faces this allows for significant improvements to face recognition technology. More details can be found in the Face Recognition section.

    Multiple Directional Illuminant Estimation from a Single Image


    We present a new method for the detection and estimation of multiple directional illuminants, using only one single image of an object of arbitrary known geometry. The surface is not assumed to be pure Lambertian, instead, it can have both Lambertian and specular properties. We propose a novel methodology that integrates information from shadows, shading and specularities in the presence of strong directional sources of illumination, even when significant non-directional sources exist in the scene. Since the specular spots have much sharper intensity changes than the Lambertian part, we can locate them in the image of the sphere by first down-sampling the image and then applying a region growing algorithm. Once the specularities have been roughly segmented, the remaining regions of the image are mostly Lambertian, and can be segmented into regions, with each region illuminated by a different set of sources, in a robust way. The regions are separated by boundaries consisting of critical points (points where one illuminant is perpendicular to the normal). Our region-based recursive least-squares method is impervious to noise and missing data. The illuminant estimation can be further refined by making use of shadow information when available, in a novel, integrated single-pass estimation. The method is generalized to objects of arbitrary known geometry, by mapping their normals to a sphere. Furthermore, we introduce a hybrid approach that combines our method with spherical harmonic representations of non-directional light sources, when such sources are present in the scene. We demonstrate experimentally the accuracy of our method, both in detecting the number of light sources and in estimating their directions, by testing on synthetic and real images.