FASTopoWM: Fast-Slow Lane Segment Topology Reasoning with Latent World Models
TLDR
FASTopoWM uses fast-slow latent world models for lane segment topology reasoning, improving temporal perception in autonomous driving.
Reasoning
The paper introduces a novel framework that leverages latent world models to enhance temporal propagation and reduce pose estimation failures, achieving state-of-the-art results on OpenLane-V2. However, the abstract is cut off and the focus is narrow (lane topology), limiting generalizability to broader world model concepts.
Read-first score
Read-first score 47.7, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 26.
Field roles
Rank sensitivity
Stability: volatile; rank range: 401.
Keyword Scores
Deep Analysis
Innovations
- Parallel supervision of historical and newly initialized queries enabling mutual reinforcement between fast and slow systems to reduce impact of pose estimation failures
- Latent query and BEV world models conditioned on action latent to propagate state representations from past observations to current timestep, improving temporal perception in the slow pipeline
Methodology
FASTopoWM is a fast-slow lane segment topology reasoning framework that uses parallel supervision of historical and newly initialized queries to mitigate pose estimation failures. It introduces latent query and BEV world models conditioned on action latent to propagate state representations temporally, enhancing the slow pipeline's temporal perception. The method is evaluated on the OpenLane-V2 benchmark.
Key Results
FASTopoWM outperforms state-of-the-art methods on OpenLane-V2, achieving 37.4% mAP (vs. 33.6%) for lane segment detection and 46.3% OLS (vs. 41.5%) for centerline perception.