Texture classification with discriminative illumination and filters. (a) Classifying aluminum and stainless steel under conventional lighting with regular color camera is challenging, since they have similar color and gloss. (b) We proposed to capture projections of BTFs for material classification with coded illumination, implemented as a LED-based multispectral dome. (c) and (d) show the optimal illumination. The bar graph shows the learned w, where the 25 bar groups correspond to the 25 LED clusters and the six bars within each group correspond to the six LEDs. This coded light pattern is also shown as w_p and w_n where w = w_p-w_n. (e) The optimal filters. (f)(g)(h) show the classification rates on test data using the VZ texture classifier [Varma05], BRDF Projection [Gu09] and our method with the same number of measurements.

The effect of orientation for BTF classification. (a) We prepared two materials coated with the same blue paint for material classification with texture only. The samples measured at different orientations show the changes in self-shadow and specular lobe caused by surface geometry. (b) By using multiple rotated samples in training, we learned classifiers (i.e., illumination and filters) that are more robust to orientation. (c) As expected, the accuracy of our BTF projection method increases with the number of rotated samples added to the training set, while the BRDF projection method [Gu09] does not vary significantly.

(a)(b) show the images of carpet and paper captured at two different scales. (c)(e): the optimal illumination (w_p, w_n) and filters trained with samples in one scale. (d)(f): the optimal illumination (w_p, w_n) and filters trained with samples in both scales. The differences in the trained illumination and filters confirm that BTF is not scale-invariant. (g): classification results when only samples in one scale are included in the training set. (h) classification results when samples in both scales are included in training set. The classification rate increases as the training sets include both scales.

The classification for aluminum and stainless steel samples. (a) Images of samples when all LEDs are turned on; (b) VZ classifier ; (c) 3D texton; (d) BRDF projection; (e) BRDF projection coupled with optimal filters; (f) Our method. The accuracy is shown in the bracket.

From top to bottom, the filter sizes are: 3 x 3, 7 x 7,11 x 11, 19 x 19 and 27 x 27, with the classification rate for the task aluminum vs. stainless shown to the bottom of each filter bank. The corresponding filters, shown in the same column, are not necessarily the scaled versions of each other due to two reasons: 1) increasing the filter size does not necessarily include more informations about the texture due to the repetition of patterns. 2) As the filter size increases, it is more likely to include outliers, such as the specular lobes, into the training set.

The filters in the optimal filter banks with different number of filters. Shown on the left side are the number of filters in the filter bank and the classification rate for the task aluminum vs. stainless steel. As shown, the performance for this task increases fast with the number of filters. This indicate that the classification of aluminum and stainless can be performed well on a subspace of BTF with lower dimensionality. The observation is similar in [Varma05] that some texture classification tasks can be performed well even though the sampling patch size is small (i.e., using more local feature). Within each filter bank, the spatial frequency of the learned filter increases with the index of filter. This indicates that the difference of the projection of BTF is concentrated on the low spatial frequencies.

This video includes more experimental results. (2.3MB)