3D shape and material properties for recognition

Description:

Shading and apparent contour as 3D cue

We will use the gauge figure method to sample local surface orientation and the correspondence method to find corresponding points in different images (e.g. due to different illumination or background) of the same object. The stimuli will be shaded objects that contain no obvious landmark attached to the object. In psychophysical experiments both photographs of real objects and computer-generated images will be used. In physiological experiments only the computer images will be used.

In a first experiment we will use pairs of images (lambertian surfaces) in which either the contour is changed but the luminance gradients remain identical (change of viewing direction) or the contour is the same but the luminance gradients are changed (change in type or direction of illumination). In the gauge experiments we will compare the pictorial relief for the two images and ascertain whether or not the location of the hyperbolic regions remain constant. In the correspondence experiments we will verify whether or not correspondence is closest (scatter minimum) for the parabolic lines (which can be inferred both from the apparent contour and the luminance gradients).

In the second experiment the surfaces will be non-lambertian with a marked highlight. Again pairs of images will be used and either direction of view or illumination will be changed. Notice that the highlight will change in either case. Both methods (gauge and correspondence) will be used as in experiment one.

In a third experiment images of single objects will be tested using the gauge method. For each point of the triangulation the depth is plotted either for two different observers, repeated testing of the same subject, or different illumination of the same object same observer. This correlation plot has either single or multiple branches, which frequently correspond to different parts of the object image. Thus multiple branches in the correlation plots, if they are consistent across the multiple replications (many pairs of subjects, many replications in a given subject, many pairs of illumination conditions in a given subject), will provide strong evidence for the presence of multiple, more or less independent parts in the mental representation of the object. The expectation is that these parts will be defined by the parabolic lines.

Material properties

In a first experiment we will use the patches of the Curet database to test whether or not human observers can judge the direction of illumination from texture. This is crucial: if the observer cannot abstract the direction of illumination, he cannot use the micro-gradients (locally) as cues for material. In the unlikely outcome that this proves true for all materials, it would become unattractive to use the Curet patches in the physiological experiments.

It is more likely that this will true for certain classes of material, depending on the accompanying illumination. In this case one should compare human judgements about direction of illumination, viewing direction and material. The experiment can be repeated with histogram-transformed textures and locally disordered textures to investigate the relative importance of the two aspects of micro-gradients.

In a second experiment we will use spherical objects and test whether or not humans can discriminate between pairs of materials. This can be tested under a variety of illumination and viewing directions and with removal of cues. First remove highlight only, then remove both highlight and microgradients.

---