Revisiting CNNs for Trajectory Similarity Learning

The paper "Revisiting CNNs for Trajectory Similarity Learning" presents ConvTraj, a novel model that uses 1D and 2D convolutions to address the limitations of RNNs and Transformers in trajectory similarity tasks. By focusing on local similarity, ConvTraj offers a simpler network structure, resulting in improved accuracy, faster training, and inference on large-scale datasets like Porto. The study compares ConvTraj favorably to Transformers and demonstrates its state-of-the-art performance in trajectory similarity search, preserving distance bounds through theoretical analysis and experimental validation. The paper also includes ablation studies, showing the importance of both convolutional layers and loss functions, and highlights the potential of CNNs for this task. The research contributes to the ongoing development of efficient and accurate trajectory analysis methods.