Research

Estimating Physical Properties from Images

In our ICCV 2015 paper we explore interactions between the appearance of an outdoor scene and the ambient temperature. We identify statistical correlations between time-lapse sequences from outdoor cameras and temperature measurements and use them to propose simple scene-specific temperature prediction algorithms which can be used to turn a camera into a crude temperature sensor.

Reflectance Fabrication and Display

We study computational techniques for fabricating and displaying user defined reflectance.

Our SIGGRAPH 2014 paper presents a reflectance display: a dynamic digital array of dots, each of which can independently display a custom, time-varying reflectance function. The display passively reacts to illumination and viewpoint changes in real-time, without any illumination-recording sensors, head tracking, or on-the-fly rendering.

In our SIGGRAPH 2013 paper we present a method to construct spatially varying reflectance at a high resolution of up to 220dpi using photolithography. An accompanying technical report explores an inexpensive alternative to micro-fabrication, in the form of metallic powders.

An Algebraic Approach to Solving Jigsaw Puzzles

We devise a new algebraic representation for edge-matching puzzles and provide conditions under which it exactly characterizes a puzzle. Using the new representation, we recast the combinatorial, discrete problem of solving puzzles as a global, polynomial system of equations with continuous variables. We further propose new algorithms for generating approximate solutions to the continuous problem by solving a sequence of convex relaxations.

Viewpoint-Aware Object Detection and Pose Estimation (editor's choice paper)

In our ICCV 2011 paper we describe an approach to category-level detection and viewpoint estimation for rigid 3D objects from single 2D images. In contrast to many existing methods, we propose a model which directly integrates 3D shape with 2D appearance. This allows us to simultaneously perform object localization and pose estimation while taking advantage of the correlation between these two tasks. In an extended version which appeared as an editor's choice paper in Image and Vision Computing we introduce a benchmark for continuous viewpoint estimation.

Cell Membrane Classification from Low Depth-Resolution EM Imagery

This project was carried out in collaboration with researchers from the Janelia Farm Campus of the HHMI as part of an effort to reconstruct the structure of massive neuronal circuits. Given only low depth-resolution imagery, we were able to accurately reconstruct cell membranes in a higher-resolution 3D space. The results are summarized in our EMMCVPR 2011 paper.

Joint Clustering of Image Segments

We developed an algorithm for segmentation of two or more closely related images, such as adjacent frames in a video sequence. In our CVPR 2011 paper we describe a functional to assess the quality of a joint segmentation by comparing two frames. The functional gives preference to selections whose shape is similar across the frames, and which are internally coherent within each frame. Our formulation allowed us to reduce the problem of finding semantically meaningful segmentations to a well-defined constrained optimization problem.

Super Resolution from a Single Image and Accurate Blur Models vs. Image Priors

We suggest an approach which combines the power of multi image, registration-based techniques with that of single image, example-based techniques in a single, unified framework. As explained in our ICCV 2009 paper, by exploiting redundancy of information in natural images, we are able to use the low-resolution input image alone, without relying on any external examples.

In our ICCV 2013 paper we examine the relative importance of the image prior and the reconstruction constraint. We further study both empirically and theoretically the sensitivity of SR algorithms to the blur model assumed in the reconstruction constraint.

Maximizing Disjoint Paths on Trees (best paper award)

In this work we present a randomized preemptive algorithm for the on-line maximum vertex disjoint paths problem on trees, and show that it has constant competitive ratio. Our result is the best possible in the sense that if one disallows either randomization or preemption, then every online algorithm cannot be better than \Omega (log n) competitive. Moreover, if the available capacity is at least 4, randomization is not needed and our online algorithm becomes deterministic. The results are described in our ALGORITHMICA paper. The preceding conference version won the best paper award at SWAT 2008.