TrAction: Action Recognition with Sparse Trajectories
Researchers have developed TrAction, a novel transformer architecture for action recognition using sparse point trajectories instead of dense video. This method aims to reduce biases found in traditional models that rely on appearance or background cues. TrAction achieves competitive accuracy on benchmarks like Something-Something V2 and EPIC-Kitchens-100, and when fused with other models, it further enhances performance. AI
IMPACT Offers a more efficient and less biased approach to video action recognition, potentially improving performance in real-world applications.