MapTCL: New training trick boosts HD map accuracy by 3.7 mAP

Online high-definition (HD) map construction is crucial for autonomous vehicles, but dynamic urban scenes with moving objects and occlusions cause geometric noise and temporal jitter between consecutive frames. Existing methods rely on feature-level temporal fusion with per-frame ground truth supervision, yet they lack an explicit objective to penalize these inconsistencies. To address this, KAIST researchers propose MapTCL, an auxiliary training strategy that introduces temporal consistency loss via bidirectional alignment between current and past frames. The method consists of two components: Bidirectional Vector Consistency Learning (BVCL), which models geometric and semantic discrepancies between associated vector instances, and Raster map Consistency Learning (RCL), which stabilizes dense BEV features.

MapTCL is a plug-and-play module that can be added to existing baseline models without any additional inference cost. On the nuScenes benchmark, it delivers gains of +3.7 mAP and +2.8 C-mAP, while on Argoverse 2 it achieves +3.1 mAP and +2.5 C-mAP. The paper, accepted at IROS 2026, demonstrates that explicitly enforcing temporal consistency during training significantly improves the reliability of generated HD maps. This makes it a practical enhancement for real-world autonomous driving systems, where map stability is essential for safe navigation.