Advances, Systems and Applications
Table 3 Our model is compared to various approaches in sport action recognition in order to determine their accuracy levels
Reference | Accuracy | Architecture | ||
|---|---|---|---|---|
UCF sport | YouTube | UCF50 | ||
Ullah et al. [13] | 96.21% | 96.40% | Optimized deep autoencoder and CNN | |
Liu et al. [61] | - | 89.7% | 93.20% | Hierarchical clustering multi-task learning |
Sadanand et al. [62] | 95.00% | - | 57.9% | High-level representation |
Tu et al. [63] | 97.53% | - | - | Multi-stream CNN |
Afza et al. [64] | 99.30% | 94.50% | - | Features fusion and weighted entropy-variances |
Muhammad et al. [65] | 99.10% | 98.30% | - | Attention based LSTM network with dilated CNN |
Meng et al. [66] | 93.20% | 89.70% | - | Spatial-temporal convolutional neural network and LSTM |
Gammulle et al. [67] | 92.20% | 89.20% | - | Two stream LSTM |
Ijjina et al. [68] | 98.90% | 94.60% | - | Hybrid deep neural network |
Zhou et al. [69] | 98.75% | 97.60% | - | Density clustering and context-guided Bi-LSTM |
Xiong et al. [70] | - | - | 96.71% | Two-Stream 3D Dilated Neural Network |
Zhang et al. [71] | - | - | 60.40% | LSTM and fully-connected LSTM with different attentions |
Dai et al. [72] | 98.90% | 96.90% | - | Two-stream attention-based LSTM |
Proposed model | 97.84% | 98.90% | 97.75% | Deformable convolution and adaptive multiscale features |