Skip to main content

Table 4 Semantic segmentation results on the S3DIS dataset (with RGB)

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

Method OA (%) AF (%) Per-class F-scores(%)
    (i) (ii) (iii) (iv) (v) (vi) (vii) (viii) (ix) (x) (xi)
Pointwise [1] 42.7 39.0 33.7 67.2 52.6 42.4 16.1 27.5 67.0 14.7 40.4 37.5 29.7
CRF-reg [22] 62.2 61.4 74.4 93.3 85.5 62.9 46.4 83.3 86.3 22.2 52.0 43.6 25.1
Seg-aided [17] 62.2 61.7 82.7 93.8 80.1 67.9 72.4 98.1 98.5 1 59.2 25.7 2
Supervised baseline 56.1 55.0 59.1 54.9 59.9 45.3 43.6 67.6 82.2 47.5 45.3 68.5 30.7
Ours no kdist nor krgb 54.1 52.3 50.5 53.1 54.6 48.0 46.6 56.7 85.6 37.6 50.5 68.1 24.1
Ours with kdist no krgb 71.9 72.2 79.4 82.1 73.6 77.3 54.0 86.9 87.4 61.4 66.4 78.0 47.5
Ours with kdist and krgb 74.5 75.1 74.2 82.3 75.7 77.6 69.2 84.2 89.7 59.0 67.5 89.8 56.6
*1 no instances of class are predicted correctly; precision=0, recall=0 - F-score undefined, taken to be 0 for the average
*2 no instances of class are predicted at all; precision undefined, recall=0 - F-score undefined, taken to be 0 for average
  1. OA overall accuracy, AF average F-score. Classes are as follows: (i) door, (ii) floor, (iii) table, (iv) window, (v) beam, (vi) book-case, (vii) ceiling, (viii) clutter, (ix) chair, (x) board, (xi) wall