TransAbs: Taming Absolute Interaction for Efficient Relative Motion Prediction

Liushuai Shi, Le Wang, Sanping Zhou, Mengnan Liu, Xiaolong Sun, Wei Tang, Gang Hua

Keywords: Absolute Position, Relative Position, Motion Prediction

Abstract: Accurate motion prediction in complex social scenarios requires capturing effective spatial interactions among traffic participants. Limited by the unsatisfactory accuracy of absolute interaction, recent methods transform the absolute scenario into multiple relative scenarios to obtain high-quality predictions. However, this relative prediction suffers from re-encoding relative motion information (spatial position and participants interaction), leading to computational inefficiency compared to absolute interaction. In this paper, we present TransAbs, which tames the absolute interaction to achieve relative effects, while maintaining the computational efficiency. The core idea of TransAbs is to map the Hadamard representation between a pair of absolute positions to their relative position. Incorporating the absolute positional embedding to the attention formulation (Hadamard product first and then sum up), TransAbs achieves relative positional embedding and spatial interaction simultaneously by leveraging a post-multiplication positional encoding. To align the Hadamard presentation and attention scores, TransAbs is optimized jointly with the motion predictor. We evaluate the effectiveness of TransAbs by integrating it into the transformer-based motion predictor commonly employed in motion prediction. Extensive experiments on a large scale public benchmark, Waymo Open Motion Dataset, demonstrate that TransAbs successfully balances prediction accuracy and computational efficiency with minimal overhead—achieving comparable accuracy while eliminating the redundant re-encoding introduced by relative interaction. The pretrained weights and code implementations will be released upon acceptance.

Primary Area: applications to robotics, autonomy, planning

Submission Number: 7706