Author(s): WENLONG, M., Institution: SHANGHAI UNIVERSITY OF SPORT, Country: CHINA, Abstract-ID: 639

INTRODUCTION:
Recognizing adversarial group activities in high-
speed sports like table tennis remains a significant challenge due
to the millisecond-scale interaction between opponents and the
complex trajectory of the ball. Traditional methods primarily
focus on collaborative semantics, which are unsuitable for captur-
ing the zero-sum dynamics inherent in competitive duels. In this
paper, we propose STAGE (Spatio-Temporal Adversarial Graph
Encoder), a novel hierarchical framework designed to explicitly
model two-player confrontational interactions. STAGE processes
information across four distinct semantic scales: Skeleton Feature
Extraction (Scale 1), Individual Action Modeling (Scale 2),
Adversarial Interaction Coupling (Scale 3), and Tactical Situation
Reasoning (Scale 4). By bridging the gap between fine-grained
joint coordinates and high-level tactical intentions, the framework
enables a panoramic analysis of match situations categorized into
proactive control, balanced confrontation, and passive constraint.
Furthermore, we introduce TTAdvantage-5K, the first dedicated
skeleton-based dataset in this domain. Extensive experiments
demonstrate that STAGE achieves state-of-the-art accuracies of
87.2% on TTAdvantage-5K and 94.3% on the VD Olympic split,
significantly outperforming existing baselines while maintaining
exceptional computational efficiency (297M FLOPs).
METHODS:
The proposed framework for analyzing two-player com-
petitive situations is organized into a hierarchical, end-to-
end pipeline. This section details the technical architecture,
beginning with a global overview of the system, followed
by the specific design of the feature acquisition modules,
the multiscale Transformer backbone, and the specialized
quantitative analysis dimensions.
RESULTS:
Results on TTAdvantage-5K. Table II presents the compar-
ison with state-of-the-art methods on TTAdvantage-5K dataset.
Our STAGE achieves the best performance with 87.2% ac-
curacy and 85.6% F1 score, significantly outperforming all
baseline methods.Results on VD Olympic Split. Table I shows the results
on VD Olympic dataset. STAGE achieves 94.3% accuracy
and 92.8% F1 score, establishing new state-of-the-art per-
formance.Efficiency Comparison. Table III compares the compu-
tational efficiency in terms of FLOPs. Traditional backbone
networks such as VGG-19 and VGG-16 require 3.6T and 2.8T
FLOPs respectively, making them less efficient for resource-
constrained scenarios.
CONCLUSION:
In this study, we presented STAGE,a Spatio-Temporal
Adversarial Graph Encoder specifically engineered to capture
the tactical dynamics of table tennis competitions. By mov-
ing beyond traditional individual action recognition, STAGE
addresses the fragmented nature of current analytics through
ahierarchical four-level semantic approach: Skeleton Feature
Extraction, Individual Action Modeling, Adversarial Interac-
tion Coupling, and Tactical Situation Reasoning.