Table 2.
Confusion matrices for the CalTech-4 dataset based on four separate category models
| Models | ||||
| Query image | F | M | A | C |
| SUVM(separate)/[Fergus et al. (14)] | ||||
| (F)ace | 0.980/0.964 | 0.069/0.33 | 0.215/0.32 | 0.252/- |
| (M)otorbike | 0.000/0.50 | 0.95/0.925 | 0.370/0.51 | 0.237/- |
| (A)irplane | 0.000/0.63 | 0.007/0.64 | 0.665/0.902 | 0.025/- |
| (C)ar | 0.000/ | 0.000/- | 0.002/- | 0.600/- |
| SUVM (shared) | ||||
| (F)ace | 0.972477 | 0.087156 | 0.674312 | 0.073394 |
| (M)otorbike | 0.007500 | 0.960000 | 0.675000 | 0.140000 |
| (A)irplane | 0.000000 | 0.002500 | 0.745000 | 0.117500 |
| (C)ar | 0.000000 | 0.000000 | 0.167500 | 0.970000 |
Each category model (j) only outputs whether a query image contains an exemplar of category (j). The table entry (i, j) is the percentage of query images belonging to category (i) that are detected to contain an instance of category (j) [by using a category model (j)]. For the top half of the table, each table entry is separated with the delimeter “/”: the number to the left is the performance of the SUVM(separate) approach and the entry to the right is the performance for Fergus et al. (14). The top half of the table corresponds to the case where SUVMs are created using separate dictionaries, whereas the bottom half of the table corresponds to the shared dictionary case. Column 1 in the top half of the table, for example, shows that the face SUVM returned no FPs when tested on nonface images; the face model in ref. 14, on the other hand, returned 50% FPs on motorbike images. Similarly, the FP rate on face images for the motorbike model is 6:9% for SUVM vs. 33% in ref. 14. The bottom half of the table shows that the face, motorbike, and car models do extremely well, even without a separate multiclass classifier.