Search

In this work, we revisit CNN architectures for sequence modeling, focusing on human trajectory prediction tasks. Forecasting human trajectories has been extensively explored as a sequence modeling problem, with many studies utilizing CNNs. Unlike conventional approaches that apply 1D convolution or 2D convolution over heatmap representations, we propose a novel architecture that applies 2D convolution directly over raw trajectory coordinates. Our method employs a coarse-to-fine strategy to refine trajectory predictions. We evaluated our approach on the ETH/UCY and Stanford Drone Datasets, demonstrating significant improvements. Our method sets a new state-of-the-art on the Stanford Drone Dataset, improving prediction accuracy and outperforming existing methods. • Unlike conventional approaches using CNNs for sequence modeling, e.g., 1D convolution or 2D convolution over the heatmap representation, we propose an architecture that applies 2D convolution over raw trajectory coordinates. • With proper padding and kernel sizes, the proposed method performs 1D convolution, 1D dilated convolution, separating coordinates convolution, and mixing features in one forward pass. • We evaluated the proposed approach on the ETH/UCY and Stanford Drone Dataset, and the proposed method improved the performance by a safe margin, setting a new state-of-the-art result on Stanford Drone Dataset. [ABSTRACT FROM AUTHOR]