Table 8 Patch-level accuracies for the stage 1 LRC on MSRC-9 for various feature sets: SIFT (S), Color (C) and Gabor (G) descriptors represented by codewords (CW) or PLSA class probabilities (CP), using only local features, or local ones + global aggregates.
From: Scene Segmentation with Low-Dimensional Semantic Representations and Conditional Random Fields
Feature set | Accuracy (%) | CPU time (s) | Dimension | |||
|---|---|---|---|---|---|---|
| Â | Local | Local + global | Local | local + global | Local | Local + global |
SC-CW | 62.2 | 75.1 | 7.7 | 197 | 1100 | 2200 |
SCG-CW | 63.2 | 77.5 | 14.1 | 1696 | 1500 | 3000 |
SC-CP | 74.7 | 76.8 | 4.8 | 6.5 | 18 | 36 |
SCG-CP | 77.1 | 79.6 | 5.8 | 8.4 | 27 | 54 |