Thomas Monninger, a recent PhD graduate from the group, will be in Arizona showcasing his works [1],[2] in WACV 2026.
AugMapNet: Autonomous vehicles need real-time, vectorized understanding of infrastructure like lanes and crosswalks. Thomas proposes AugMapNet, which utilizes a novel latent Bird’s-Eye View grid augmentation technique. By better structuring the latent space and combining vector decoding with dense spatial supervision, they achieve 13.3% performance improvement over baselines on the nuScenes dataset.
NavMapFusion: Thomas introduces NavMapFusion, a generative online mapping approach that uses widely available Standard-Definition maps as coarse priors. The model treats potential discrepancies between the map and real-time sensor data as noise in the diffusion process, yielding a 21.4% improvement in map construction accuracy.
[1] T. Monninger, M. Z. Anwar, S. Antol, S. Staab, S. Ding. AugMapNet: Improving Spatial Latent Structure via BEV Grid Augmentation for Enhanced Vectorized Online HD Map Construction. In: The IEEE/CVF Winter Conference on Applications of Computer Vision (WACV 2026), Tucson, Arizona, March 6-10, 2026.
[2] T. Monninger, Z. Zhang, S. Staab, S. Ding. NavMapFusion: Diffusion-based Fusion of Navigation Maps for Online Vectorized HD Map Construction. In: The IEEE/CVF Winter Conference on Applications of Computer Vision (WACV 2026), Tucson, Arizona, March 6-10, 2026.