21 April 2021 in IEEE Access
This paper addresses the problem of event detection and localization in long football (soccer) videos. Our key idea is that understanding the long-range dependencies between video frames is imperative for accurate event localization in long football videos. Additionally, proper event detection is not likely for fast movements in football videos without considering mid-range and short-range correlations between neighboring video frames. We argue that event spotting can be considerably improved by considering short-range to long-range frame dependencies in a unified architecture. To model long-range and mid-range dependencies, we propose to use the dilated recurrent neural network (DilatedRNN) with long short-term memory (LSTM) units, grounded on two-stream convolutional neural network (Two-stream CNN) features. While two-stream CNN extracts local spatiotemporal features necessary for fine-level details, the DilatedRNN makes the information obtained from distant frames available for the classifier and spotting algorithms. Evaluating our event spotting algorithm on the largest publicly available benchmark football dataset -SoccerNet- shows an accuracy improvement of 0.8% - 13.6% compared to state of the art, and up to 30.1% accuracy gain in comparison to the baselines. We also investigate the contribution of each neural network component in spotting accuracy through an extensive ablation study.
1 January 2025 in Empathic Computing
Multimodal emotion recognition with disentangled representations: private-shared multimodal variational autoencoder and long short-term memory framework
2 July 2026 in Empathic Computing
Aims: Multimodal emotion recognition (MER) can outperform unimodal approaches by integrating complementary information from multiple sources. However, real-world applications often involve incomplete or missing modalities, reducing the reliability of existing MER models. This study proposes a framework that remains robust under missing-modality conditions while preserving the advantages of multimodal integration. Methods: To address this challenge, we propose a cross-modal latent regeneration and attention long short-term memory (CMLR-ALSTM) framework. The framework combines pretrained variational autoencoder encoders with residual projection networks, optimized with L2 loss, to achieve stable latent-space alignment across modalities. Regenerated and available latent representations are integrated using a cross-modal attention mechanism and processed by an LSTM to model temporal dependencies and improve multimodal fusion under incomplete data. Results: The proposed framework was evaluated on three benchmark datasets under complete, partially missing, and completely missing modality scenarios. Experimental results demonstrate that CMLR-ALSTM achieves up to 17.22% improvement under missing-modality conditions while maintaining competitive performance under complete-modality settings. The results also show that the proposed latent regeneration strategy effectively preserves cross-modal relationships and robust latent representations. Conclusion: The experimental results confirm the effectiveness of the proposed framework, particularly in realistic environments where data availability is inconsistent. By leveraging CMLR to reconstruct missing representations and modelling temporal dependencies through LSTM, the proposed approach provides a more robust and reliable MER framework for practical deployment. The results across different modality combinations demonstrate its ability to generalize across heterogeneous multimodal settings.