Symphonies is a novel approach to 3D Semantic Scene Completion (SSC), a task that is becoming increasingly important for autonomous driving. The method focuses on completing the scene volume from a sparse set of instance queries derived from the input, with an emphasis on context awareness. This approach allows Symphonies to dynamically encode instance-centric semantics, interacting with image and volume features while avoiding the need for dense voxel-wise modeling. This method also captures context throughout the entire scene, which helps to alleviate geometric ambiguity that can arise from occlusion and perspective errors. The paper demonstrates that Symphonies achieves state-of-the-art results on the challenging SemanticKITTI dataset, outperforming existing methods and showcasing the promising advancements of the paradigm.
Publication date: June 27, 2023
Project Page: https://github.com/hustvl/Symphonies
Paper: https://arxiv.org/pdf/2306.15670.pdf
