Table 1.
Model | M1 | M1A | M1B | |||
---|---|---|---|---|---|---|
Noise | σ = 0 | σ = 0.3 | σ = 0 | σ = 0.3 | σ = 0 | σ = 0.3 |
Precision | 97.2 % | 96.2 % | 97.2 % | 97.2 % | 86.7 % | 82.1 % |
Recall | 95.7 % | 93.1 % | 96.4 % | 85.9 % | 86.4 % | 84.3 % |
F-score | 96.5 % | 94.6 % | 96.8 % | 91.2 % | 86.5 % | 83.1 % |
MAE | 0.2 | 0.2 | 0.2 | 0.5 | 0.4 | 0.5 |
MAHD | 2.3 px | 3.1 px | 2.0 px | 5.0 px | 4.0 px | 6.6 px |
MAHD⋆ | 1.4 px | 1.8 px | 1.4 px | 3.1 px | 2.4 px | 3.6 px |
PS predictions were performed from 5 excitation frames and are shown in Figure 9. MAE is the mean absolute error of the number of predicted PS, MAHD is the mean average Hausdorff distance. The test dataset contains 5, 000 frames with 17, 360 PS in total, however 420 frames contain no ground truth PS. If the model predicts any PS for a frame which contains no ground truth PS—or if no PS are predicted for a sample which does contain ground truth PS—we assign a maximum average Hausdorff distance (181 px) for the computation of the MAHD. This skews the MAHD significantly. Accordingly, MAHD⋆ is the MAHD when we ignore these samples.