A Method For Group Activity Recognition In Volleyball Videos With Extensions To Domain Generalization

Official URL: https://risc01.sabanciuniv.edu/record=b3205815

In this thesis, we present two novel methods to address the challenges of group activity recognition and domain generalization: DECOMPL and ADRMX, respectively. Our primary focus is on the recognition of group activities in volleyball videos. We argue that previous temporal methods have not shown significant performance improvements that justify their additional computational cost, which scales linearly with the number of frames. To tackle this, we propose DECOMPL, a non-temporal method that leverages both visual and coordinate features from a single frame to classify the activity in a video. For the task of group activity recognition in volleyball videos, we introduce several problem-specific contributions. These include utilizing horizontal flips to exploit the symmetry of activities, decomposing labels to provide additional feedback through subtasks, and employing a heuristic to split team features. Furthermore, during our study of the Volleyball dataset, which is widely used in recent literature, we realized that the labeling scheme degrades the group concept, reducing them to the level of individual actions. We correct for this by providing new reannoations that emphasize the group concept. DECOMPL demonstrates remarkable performance on both the Volleyball dataset and the Collective Activity dataset, showcasing its effectiveness in group activity recognition. Our approach is on par with temporal methods, highlighting its potential in this field. In addition to group activity recognition, we also investigate the domain generalization problem, as videos often come from different domains due to variations in camera orientation and background or due to even the team side change in volleyball videos. ADRMX, our proposed method for domain generalization, incorporates domain variant features along with domain invariant ones with an additive disentanglement. To enhance the robustness of our model, we introduce a novel data augmentation technique called remix strategy, which operates on the latent space to generate synthetic instances. On the DomainBed benchmark, ADRMX achieves state-of-the-art performance among 14 algorithms, as measured by average accuracy across seven well-known datasets.