Abstract
The estimation of 3D shape from 2D images requires processing and combining many cues, including texture, shading, specular highlights and reflections. Previous research has shown that oriented filter responses ('orientation fields') may be used to perceptually reconstruct the surface structure of textured and shaded 3D objects. However, texture and shading provide fundamentally different information about 3D shape -- texture provides information about surface orientations (which depend on the first derivative of the surface depth) while shading provides information about surface curvatures (which depend on higher derivatives). In this research project, we used specific geometric transformations that preserve the informativeness of one cue's orientation fields while disturbing the other cue’s orientation fields to investigate whether oriented filter responses predict the observed strengths and weaknesses of texture and shading cues for 3D shape perception.
In the first experiment, a 3D object was matched to two comparison objects, one with identical geometry and another with a subtly different pattern of surface curvatures. This transformation alters second derivatives of the surface while preserving first derivatives, so changes in the orientation fields predict higher detectability for shaded objects. This was reflected in participants' judgments and model performance. In the second experiment, observers matched the perceived shear of two objects. This transformation alters first derivatives but preserves second derivatives. Therefore, changes in the orientation fields predicted a stronger effect on the perceived shape for textured objects, which was reflected in participant and model performance.
These results support a common front-end -- based on orientation fields -- that accounts for the complementary strengths and weaknesses of texture and shading cues. Neither cue is fully diagnostic of object shape under all circumstances, and neither cue is 'better' than the other in all situations. Instead, the model provides a unified account of the conditions in which cues succeed and fail.
Meeting abstract presented at VSS 2013