My research interests lie at the rich intersection of human visual psychophysics, computer vision and machine learning. I am interested in problems in visual perception both from a behavioral (what can humans perceive) and a computational (how are visual stimuli processed in the brain or a computer) perspective.

In my thesis work, I have focused on the problem of human material perception, an emerging field of study. We interact with a variety of materials in daily life e.g. fabric, paper, metal, plastic etc. We constantly make judgments of material appearance such as is that patch of pavement icy, are those fruits ripe, is that bag made of real leather, or is that shirt clean? In spite of the ubiquity of material judgments, little is known about how material perception is achieved.

Examples of Material Judgments

Along with my collaborators, I have studied different aspects of human material perception. In Sharan et al. 2008 and Motoyoshi et al. 2007, we demonstrated that the perception of surface albedo and gloss for some natural surfaces is correlated with certain image statistics like moments and percentiles of the luminance histogram and filter outputs. In ongoing research, we show that material perception on a diverse set of natural images is fast, rich and flexible. Human observers can perceive a variety of material properties in brief presentations (40 msec followed by mask). This performance cannot be explained either by low-level cues (e.g color, resolution) or high-level information like object identity.

Please visit the individual project pages for more details (click on image or project title).

 

  

Oreo Cookie made of knit wool

Fake food, an Oreo Cookie made of knit wool. Observers can identify both the object and the material in 40 msec viewing.

Rapid perception of Material Properties in Natural Images

Lavanya Sharan, Ruth Rosenholtz & Edward H. Adelson

We studied material judgments for a wide range of images of complex, real world materials. It is well known that observers can quickly extract a lot of information about objects and scenes from photographs that they have never seen before (Biederman et al. 1974; Potter 1975, 1976; Thorpe et al. 1996). In a set of five experiments, we show that material perception is fast and flexible, and can have the same rapidity as object recognition and scene perception. For example, even in a 40 msec presentation followed by masking, observers can tell that the image on the left is that of an Oreo cookie made of knit wool.

 

 

Two version of Michelangelo

Consider these two renderings of Michaelangelo's sculpture of St. Matthew (image courtesy Digital Michelangelo Project). The one on the left looks glossier and darker than the one on the right. However, both images have the same mean luminance i.e. the same average amount of light reaches your eyes from each image. Previous work on surface appearance has focused only on mean luminance, and therefore cannot explain this effect.

 

Image Statistics & the Perception of Albedo and Gloss

Lavanya Sharan, Yuanzhen Li, Isamu Motoyoshi, Shin'ya Nishida, Edward H. Adelson

In our papers, we show that certain image statistics like moments (standard deviation, skewness) and percentiles (10th, 90th) of the luminance histogram and filter outputs are correlated with surface gloss and surface albedo for certain natural surfaces. These statistics are correlated not only with the physical reflectance parameters, but also with human judgments of surface reflectance. Manipulating these statistics in an image alters the perception of reflectance. How might the human brain compute these statistics? For skewness, we offer an explanation in terms of ON and OFF center filtering mechanisms that are well known in the early visual pathway. In order to test for the existence of skewness detection mechanisms in the brain, we developed a novel visual aftereffect.

Papers: Motoyoshi, Nishida, Sharan & Adelson, Image statistics and the perception of surface qualities, Nature, 2007. Supplementary material

Sharan, Li, Motoyoshi, Nishida & Adelson, Image statistics for surface reflectance perception, JOSA A, 2008.

 

Doll HDR Dataset

Tone-mapped output of our algorithm on the Doll HDR Dataset.

Compressing & Companding High Dynamic Range Images

Yuanzhen Li, Lavanya Sharan & Edward H. Adelson

High dynamic range (HDR) imaging is an area of increasing importance, but most display devices still have limited dynamic range (LDR). Various techniques have been proposed for compressing the dynamic range while retaining important visual information. Multiscale image processing techniques, which are widely used for many image processing tasks, have a reputation of causing halo artifacts when used for range compressing. However, we demonstrate that they can work when properly implemented. We use a symmetrical analysis-synthesis filter bank, and apply local gain control to the subbands. We also show that the technique can be adapted for the related problem of "companding", in which an HDR image is converted to an LDR image, and later expanded back to high dynamic range.

Paper: Li, Sharan & Adelson, Compressing and companding high dynamic range images with subband architectures, SIGGRAPH 2005.