Search papers, labs, and topics across Lattice.
TopoMaskV3 introduces a 3D mask head for road topology understanding, addressing limitations of prior 2D mask-based methods by incorporating dense offset and height predictions to mitigate discretization artifacts. This approach enables direct 3D estimation from a rasterized representation, eliminating the need for fusion with parametric heads. The paper also pioneers geographically distinct splits and a long-range benchmark to address data leakage, achieving state-of-the-art performance with 28.5 OLS on the new benchmark.
TopoMaskV3's 3D mask head leapfrogs prior 2D methods for road topology understanding, proving that dense offset and height predictions can conquer discretization artifacts and geographic overfitting.
Mask-based paradigms for road topology understanding, such as TopoMaskV2, offer a complementary alternative to query-based methods by generating centerlines via a dense rasterized intermediate representation. However, prior work was limited to 2D predictions and suffered from severe discretization artifacts, necessitating fusion with parametric heads. We introduce TopoMaskV3, which advances this pipeline into a robust, standalone 3D predictor via two novel dense prediction heads: a dense offset field for sub-grid discretization correction within the existing BEV resolution, and a dense height map for direct 3D estimation. Beyond the architecture, we are the first to address geographic data leakage in road topology evaluation by introducing (1) geographically distinct splits to prevent memorization and ensure fair generalization, and (2) a long-range (+/-100 m) benchmark. TopoMaskV3 achieves state-of-the-art 28.5 OLS on this geographically disjoint benchmark, surpassing all prior methods. Our analysis shows that the mask representation is more robust to geographic overfitting than Bezier, while LiDAR fusion is most beneficial at long range and exhibits larger relative gains on the overlapping original split, suggesting overlap-induced memorization effects.
