Details
- Supervisors
- Faculty
- Degree label
- Abstract
- Template-based gesture recognition remains a widely used approach due to its simplicity, low computational cost, and absence of a training phase. However, individual recognizers often exhibit uneven performance across gesture classes and datasets, which limits their robustness. Ensemble learning offers a potential solution by combining multiple recognizers in order to exploit complementary strengths and mitigate individual weaknesses. This thesis investigates whether ensemble learning can improve recognition accuracy and robustness when applied to template-based gesture recognizers. Several ensemble strategies are implemented and evaluated within the QuantumLeap framework, including voting-based methods, stacking with a meta-learner, and a Dirichlet-based weighted averaging approach. The study focuses on three datasets of increasing difficulty: SHREC2019, 3DTCGS, and Kinemic, allowing ensemble behavior to be analyzed under diverse conditions. Experimental results show that ensemble learning consistently outperforms the average accuracy of the composing recognizers, with only rare exceptions. Improvements are particularly significant on challenging datasets where individual recognizers perform poorly and leave more room for enhancement. However, ensembles do not systematically surpass the best individual recognizer, especially when performance is already close to a ceiling. A strong relationship is observed between ensemble effectiveness and diversity among recognizers, measured through variance in their predictions. Among the evaluated approaches, the Dirichlet-based ensemble with statistically derived weights achieves the highest peak performance and the most stable improvements. These findings demonstrate that ensemble learning is a valuable strategy for template-based gesture recognition, particularly when robustness and consistency are prioritized over marginal gains beyond the best individual recognizer.