SiFH: Siamese frequency harmonization self-supervised learning for motion forecasting

Chunyu Liu

doi:10.1016/j.neucom.2025.131563

SiFH: Siamese frequency harmonization self-supervised learning for motion forecasting

2025-09-15

0

PRE

AI

DOI：10.1016/j.neucom.2025.131563

Original

Original request for help

Save

Abstract

En 中文

Motion forecasting is a fundamental component of the autonomous driving system and plays an important role in ensuring safety. While supervised learning methods have achieved promising performance in many domains, their model capacity is typically limited by the availability of annotated data. Some previous works have tried to introduce this paradigm into motion forecasting. However, these early studies neglect the task-specific characteristics. To address this, we incorporate two key insights into motion forecasting tasks within the self-supervised paradigm and propose a novel self-supervised motion forecasting framework. First, we design a Frequency Information Harmonization pretext task that explicitly encourages the model to integrate frequency domain features with their time domain counterparts, making them work harmoniously. Second, we introduce an Implicit Scene Alignment task, which enables the model to learn scene-level semantics by aligning masked and unmasked views through shared prototypes. By jointly optimizing these objectives, the model is encouraged to leverage abundant unlabeled data and capture rich spatio-temporal representations. Extensive experiments conducted on the challenging Argoverse 2 and Argoverse 1 benchmarks demonstrate that our proposed model outperforms previous state-of-the-art baselines and can produce more accurate and reliable predictions.

SiFH: Siamese frequency harmonization self-supervised learning for motion forecasting

Abstract

Journal

Organization