Kernkonzepte
The proposed Cross-Block Fine-Grained Semantic Cascade (CFSC) module effectively captures and aggregates low-level fine-grained features from shallow blocks of the GCN backbone to improve the classification accuracy of sports actions.
Zusammenfassung
The paper presents a novel "Cross-Block Fine-Grained Semantic Cascade (CFSC)" module to address the challenge of fine-grained sports action recognition. Existing GCN-based methods often focus on constructing generalized topology graphs and aggregating features within blocks, lacking the utilization of low-level detail features and cross-block feature fusion.
The CFSC module works as follows:
- It extracts feature maps from multiple depth blocks of the GCN backbone, capturing features of varying granularity.
- It applies temporal convolution on the features at each level to learn short-term temporal dependencies.
- It progressively aggregates the features from shallow to deep blocks using element-wise addition, allowing the network to focus on action details.
- The aggregated highest-level feature is then normalized and passed through ReLU to generate the auxiliary discriminative feature.
This cross-block feature aggregation methodology helps mitigate the loss of fine-grained information, leading to improved performance on sports action classification tasks.
The authors also collected a new fencing dataset, FD-7, with 7 classes of high-speed fencing actions. Experiments on FD-7 and the public FSD-10 figure skating dataset demonstrate the advantage of the CFSC module over state-of-the-art methods.
Statistiken
The velocity at the tip of the sword in fencing can reach up to 150 meters per second.
The rotational speed of figure skaters can reach up to six revolutions per second following a jump.
Zitate
"Contemporary studies have provided compelling evidence that convolutional layers at different depths encompass information across diverse granularities."
"Skeleton-based methods can effectively capture the intrinsic characteristics of motions and exhibit strong robustness."