TITLE:
Investigating the Effect of Temporal Sequence Length and Additive Attention Pooling on BiLSTM-Based Fencing Outcome Recognition
AUTHORS:
Qiuhong Wu, Yang Zhang
KEYWORDS:
Fencing Outcome Recognition, Human Pose Estimation, OpenPose, BiLSTM, Temporal Sequence Length, Additive Temporal Attention, Imbalanced Classification
JOURNAL NAME:
Journal of Computer and Communications,
Vol.14 No.10,
October
22,
2026
ABSTRACT: We examine the impacts of temporal sequence length and additive temporal attention pooling on the automated classification performance of a Bidirectional Long Short-Term Memory (BiLSTM) model for fencing outcome recognition. Human pose data was modelled via the OpenPose BODY_25 keypoints, and the fencing outcomes were then labeled in each exchange (i.e., left-fencer score, right-fencer score, no valid score). A controlled 3 × 2 experimental design was employed with varying sequence lengths (25, 50, or 75 frames) with or without additive temporal attention. All six setups were evaluated on a single fixed source-disjoint training, validation and test split with five matched training seeds for each setup. Only the best validation accuracy was used for model selection. Our best performing model was an attention model with a sequence length of 25 frames and a mean best validation accuracy of 58.93%. Nevertheless, its 0.18-percentage-point improvement on the 25-frame no-attention model was both statistically and practically insignificant. For no-attention, increasing the sequence length did not yield meaningful performance gain. We see that introducing attention did not make much of a difference when we used 25 frames, but improved the mean best validation accuracy by 1.77 percentage points when we used 50 frames, and 3.23 percentage points when we used 75 frames, when compared to the no-attention version of the respective model. The test accuracy of the model, which was chosen at the best validation performance, was 57.35%, with a macro-F1 of 48.29% and a balanced accuracy of 51.14%. Performance on the underrepresented class of no valid score remained poor. This study suggests that we may not benefit from longer temporal windowing for fencing outcome recognition and additive temporal attention may become counterproductive at longer sequencing. The proposed classifier, hence, holds its value as an experimental model but is not yet effective as a refereeing tool.