Table 2.
Average confusion matrices for six types of features over 10 iterations of 3-fold cross-validation for a codebook of size 40 clusters. Each row shows the true class labels and columns show predicted labels. Each row sums to the total number of images in that subtype.
| (a) No Inv. | ||||
|---|---|---|---|---|
| CC | CH | ON | PA | |
| CC | 19.5 | 8.1 | 0.9 | 0.5 |
| CH | 10.0 | 18.2 | 2.9 | 0.9 |
| ON | 1.8 | 5.9 | 8.5 | 0.8 |
| PA | 1.5 | 1.2 | 1.5 | 23.8 |
| (b) No Inv. (dense) | ||||
|---|---|---|---|---|
| CC | CH | ON | PA | |
| CC | 18.3 | 8.6 | 1.6 | 0.5 |
| CH | 8.4 | 18.9 | 3.0 | 1.7 |
| ON | 3.3 | 2.5 | 9.3 | 1.9 |
| PA | 3.6 | 3.2 | 2.8 | 18.4 |
| (c) Scale Inv. | ||||
|---|---|---|---|---|
| CC | CH | ON | PA | |
| CC | 19.5 | 8.4 | 1.1 | 0.0 |
| CH | 8.6 | 19.2 | 3.0 | 1.2 |
| ON | 2.0 | 4.4 | 9.6 | 1.0 |
| PA | 1.3 | 1.0 | 1.9 | 23.8 |
| (d) Rot Inv. | ||||
|---|---|---|---|---|
| CC | CH | ON | PA | |
| CC | 25.0 | 2.9 | 0.7 | 0.4 |
| CH | 5.5 | 25.4 | 1.1 | 0.0 |
| ON | 1.7 | 2.9 | 11.0 | 1.4 |
| PA | 0.7 | 0.1 | 1.8 | 25.4 |
| (e) Rot. Inv. (dense) | ||||
|---|---|---|---|---|
| CC | CH | ON | PA | |
| CC | 25.7 | 1.9 | 1.0 | 0.4 |
| CH | 4.3 | 25.6 | 1.7 | 0.4 |
| ON | 1.4 | 3.0 | 11.4 | 1.2 |
| PA | 1.1 | 0.7 | 2.5 | 23.7 |
| (f) Scale & Rot. Inv. | ||||
|---|---|---|---|---|
| CC | CH | ON | PA | |
| CC | 25.3 | 3.3 | 0.1 | 0.3 |
| CH | 3.5 | 25.9 | 2.1 | 0.5 |
| ON | 0.8 | 1.8 | 13.8 | 0.6 |
| PA | 1.4 | 0.3 | 1.3 | 25.0 |