Skip to main content
. Author manuscript; available in PMC: 2022 Nov 5.
Published in final edited form as: Med Image Comput Comput Assist Interv. 2022 Sep 16;13438:130–139. doi: 10.1007/978-3-031-16452-1_13

Table 1.

Comparison with baseline methods. Performance is evaluated using macro F1 score, precision, and recall. We find that pre-training results in significantly improved performance over training from scratch and the other methods.

Method F1 Pre Rec
GaitForeMer (Ours) 0.76 0.79 0.75
GaitForeMer-Scratch (Ours) 0.60 0.64 0.58
OF-DDNet♮* [15] 0.58 0.59 0.58
ST-GCN [26]* 0.52 0.55 0.52
DeepRank♮* [18] 0.56 0.53 0.58
SVM♮* [24] 0.44 0.49 0.40
♮

refers to results directly cited from [15].

*

indicates statistical difference at (p < 0.05) compared with our method, measured by the Wilcoxon signed rank test [25]. Note that this is a 4-class classification problem and hence 0.25 recall implies a random classifier. Best results are in bold. See text for details about compared methods.