Table 7.
The F-ratio quantifies the variance in word meaning captured by the models. Our CIRCA method produced a superior fit to labelling data when we used the optimal number of clusters (24 and 55), but performed worse when we matched the number of clusters in SUN.
| Model | No. of Images | K categories | F |
|---|---|---|---|
| CIRCA | 1,000 | 55 | 537.22 |
| SUN | 1,000 | 72 | 447.36 |
| CIRCA | 1,000 | 72 | 417.68 |
| CIRCA | 712 | 24 | 1127.15 |
| Greene | 712 | 22 | 1121.06 |
| CIRCA | 712 | 22 | 1190.57 |
| SUN | 712 | 35 | 815.49 |
| CIRCA | 712 | 35 | 794.52 |