Skip to main content

Table 7 Parameter studies on varying the number of initially annotated points, tested on the Oakland dataset

From: Pseudo-labelling-aided semantic segmentation on sparsely annotated 3D point clouds

Method Overall accuracy (%) Average F-score (%) Per-class F-scores(%)
    Foliage Wire Pole Ground Façade
5 labels/class
Seg-aided [17] 95.1 58.9 95.1 4.1 14.3 93.9 87.1
Ours with kdist 80.6 45.2 0.0 33.9 32.7 99.4 60.2
15 labels/class
Seg-aided [17] 96.6 68.4 93.7 46.5 8.7 99.5 93.9
Ours with kdist 96.6 74.2 92.0 40.2 46.2 99.3 93.3
30 labels/class
Seg-aided [17] 96.2 67.8 94.8 33.6 19.3 98.9 92.2
Ours with kdist 95.9 70.2 90.5 34.1 35.5 99.3 91.4
100 labels/class
Seg-aided [17] 96.2 68.0 95.5 32.2 20.7 98.8 92.4
Ours with kdist 96.2 72.6 91.4 38.3 41.4 99.4 92.4
  1. Experiments were performed using our method and segmentation-aided classification