Beyond Vision: A Semantic Reasoning Enhanced Model for Gesture Recognition with Improved Spatiotemporal Capacity

Abstract

AbstractGesture recognition is an imperative and practical problem owing to its great application potential. Although recent works have made great progress in this field, there also exist three non-negligible problems: 1) existing works lack efficient temporal modeling ability; 2) existing works lack effective spatial attention capacity; 3) most works only focus on the visual information, without considering the semantic relationship between different classes. To tackle the first problem, we propose a Long and Short-term Temporal Shift Module (LS-TSM). It extends the original TSM and expands the step size of shift operation to model long-term and short-term temporal information simultaneously. For the second problem, we expect to focus on the spatial area where the change of hand mainly occurs. Therefore, we propose a Spatial Attention Module (SAM) which utilizes the RGB difference between frames to get a spatial attention mask to assign different weights to different spatial positions. As for the last, we propose a Label Relation Module (LRM) which can take full advantage of the relationship among classes based on their labels’ semantic information. With the proposed modules, our work achieves the state-of-the-art performance on two commonly used gesture datasets, i.e., EgoGesture and NVGesture datasets. Extensive experiments demonstrate the effectiveness of our proposed modules.KeywordsGesture recognitionTemporal modelingSpatial attentionSemantic relation