Table 2.
Comparison of design choices for two-stream clustering. We compute the object detection accuracy by assigning the labels of the cluster centers to other cluster members. The number of candidates per cluster, Q, is fixed to 5.
Description | Random | One-Stream |
Two-Stream |
||||||||
---|---|---|---|---|---|---|---|---|---|---|---|
Mask-Only | Image-Only | Late-Fusion | Hierarchical | ||||||||
Es clusters (N) | - | - | 128 | 256 | 1 | 1 | - | 64 | 128 | 64 | 32 |
Eu clusters (M) | - | - | 1 | 1 | 128 | 256 | - | 2 | 2 | 4 | 8 |
Total num. (MN) | - | - | 128 | 256 | 128 | 256 | 256 | 128 | 256 | 256 | 256 |
Annotation ratio (%) | 2.23 | 4.46 | 2.23 | 4.46 | 2.23 | 4.46 | 4.46 | 2.23 | 4.46 | 4.46 | 4.46 |
Accuracy | 0.767 | 0.772 | 0.805 | 0.819 | 0.420 | 0.578 | 0.738 | 0.821 | 0.826 | 0.846 | 0.814 |