Abstract
Many lightness experiments and illusions suggest that mid-level features such as lighting boundaries and reflectance edges play an important role in lightness perception. However, there has been relatively little work on developing complete computational models of lightness that incorporate mid-level factors, compared to the extensive work on developing low-level computational models. Here I use probabilistic graphical models and their well-developed inference methods to formulate a mid-level computational model of human perception of lightness and lighting. To simplify a first approach, I model lightness perception on a 16 × 16 pixel grid. Within this domain one can create many lightness illusions (e.g., the argyle illusion) and phenomena (e.g., assimilation) that challenge current models. The model makes simple probabilistic assumptions about local properties of lighting and reflectance:(1) reflectance spans the range 3% to 90%, (2) reflectance tends to change slowly from place to place, (3) incident illuminance spans the range 0 to 100,000 lux, (4) illuminance edges are less common than reflectance edges, (5) illuminance edges tend to be straighter than reflectance edges, (6) reflectance and illuminance edges usually occur at image image luminance edges, and (7) measured image luminance is contaminated by a small amount of noise. The model uses these local assumptions along with belief propagation methods to infer globally optimal estimates of reflectance and illuminance in 16 × 16 pixel stimulus images. The model arrives at human-like interpretations of several lightness illusions that have been problematic for previous models, including the argyle illusion, snake illusion, Koffka ring, and their control conditions. The model also reproduces several important lightness phenomena, including contrast and assimilation effects. Thus a probabilistic graphical model that incorporates simple assumptions about reflectance and lighting provides a strong mid-level computational account of lightness perception over a wide range of stimulus conditions.
Acknowledgement: NSERC, NVIDIA