Fig. 1. Illustrations of mis-segmentation and mis-labeling. We represent blobs segmented with Normalized Cuts [1] and [6]. On the left, the penguin is split into five blobs, moreover, like in the right illustration, we do not know the relation between labels and blobs.

Fig. 2. Words distributions over A(w_k), B(w_k) and c_T/c_G showing that C(x;w_k) is negligible in estimations for nearly all the words. These distributions have been calculated on 6K images (around 60K blobs), considering that for each word w_k, C_S equals the number of images labeled by w_k.

Fig. 3. Maximum values of normalized estimated discriminant power , in each SHAPE, COLOR or TEXTURE feature set. Each word is represented by a dot followed by the most frequent word, possibly indexed by the dimension index of the shape feature maximizing : 4 is the ratio of blob's area to the perimeter squared; 6 the ratio of the blob's area by its convex hull. We draw the equal line of discriminant power for easy comparison. Results are intuitively correct: FLOWER, SNOW, PLANT are better discriminated by color; LEAF by color and shape; WATER and TREE (variant shapes and colors) by texture. Ratio of the blob's area by its convex hull is nearly always more discriminant for FLAG than any color or texture features. DOOR and ENTRANCE are more discriminated by shape than by color.

Fig. 4. Example of images labeled by words which have a high shape feature discrimination power: (top to bottom) HORIZON, BIGHORN, FLAG, DOOR. One can see indeed that these images contain recurrent shape, but not recurrent color neither texture, yielding to highest shape feature discrimination power.

Overview of classification results (mean of Normalized Score (NS)) under baseline AHC system with 40 features (40DIM), or with using only selected features with ALDA (NADAPT) for τ=0.3 (Eq. (9))

NS are averaged over the 52 most frequent words and 2.5K COREL test images.