WorldSculpt：基于单物体生成先验解构复杂三维场景

Translation Task Breakdown

Okay, so the task is clear. I need to translate this abstract into simplified Chinese. It's a technical text from the 3D generation field, focusing on a compositional 3D representation of cluttered scenes. I understand the nuances of this area, so I'll need to maintain accuracy and precision in the terminology. Paragraph breaks are to be preserved, and there are no URLs to worry about, just a straight translation. I've broken down my approach. First, I have to analyze the source.

The first step is to dive into the core concepts. The abstract deals with generating a compositional 3D representation of a scene with many objects. The goal is to represent the scene as a collection of meshes, which is useful for applications such as gaming, AR/VR, simulation, and robotics. This becomes complicated in cluttered scenes where objects are heavily occluded. Geometry-based approaches are limited because of the reconstruction as a single representation, while compositional methods are limited to simple scenes. So, the authors present a method of generating complex scenes compositionally by adapting a single-object 3D generative prior to multi-view observations. Pixal3D implements this, extending it with a multi-view conditioning pathway. The model generalizes to large scenes with severe occlusion. They've also introduced UE-MeshyScene, a photorealistic benchmark of densely cluttered scenes with many objects. This new method outperforms existing ones, and they also show how to convert generated 3DGS worlds into compositional mesh scenes.

I need to nail down the terminology and style. I need to get the "feel" right. "Compositional 3D representation" requires a direct and clear translation. "Cluttered scene" needs careful handling. "Shared world frame" should be "统一世界坐标系". "Individual object meshes" should be translated as "独立物体网格". "Canonical space" is "规范空间". "Posed observations" are best rendered as "带有位姿的观测". "3DGS worlds" translates into "3DGS 世界", and "ground-truth meshes" becomes "真值网格". I also have to use “将物体生成锚定（建立）在……之上” instead of “扎根在……之中”.

Now, I'm ready to write the translation.

我们研究了为包含数百个物体的杂乱场景生成组合式3D表征的问题。其目标是将场景表示为放置在统一世界坐标系下的独立物体网格的集合，以满足游戏、AR/VR、仿真和机器人等下游应用的需求。在密集杂乱的场景中，该任务极具挑战性，因为物体之间存在严重的相互遮挡，且每个视角仅能展现其几何结构的一小部分。基于几何的方法通常将场景重建为单一表征，并在遮挡区域留下不完整的几何形态；而现有的结合生成先验的组合式方法则大多局限于相对简单的场景。我们表明，通过将强大的单物体3D生成先验适配到多视角观测中，能够以组合式方式生成包含数百个物体的复杂场景。我们通过 Pixal3D 实现了这一范式，并通过扩展多视角条件路径，将物体生成扎根于多个带位姿的观测之中。尽管该模型完全是在规范空间的单个物体上微调的，但在没有任何场景级训练的情况下，它仍能泛化到具有严重遮挡的大规模场景中，这证明了该范式的可行性与可扩展性。我们进一步推出了 UE-MeshyScene，这是一个包含数百个物体的密集杂乱场景的照片级真实感基准，提供了逐物体的标注以及真值网格。在单物体、受控多物体以及 UE-MeshyScene 的各项评测中，我们的方法均持续优于先前的方法，且随着场景复杂度和遮挡程度的增加，优势更为显著。最后，我们通过将生成的 3DGS 世界（例如 Marble 和 HY-World 2.0）转换为组合式网格场景，展示了更广泛的应用前景。

Finally, I'll make sure I've met all the constraints. Simplified Chinese? Check. Paragraph breaks preserved? Check. No URLs? Check. Just the translation? Check. Ready to go.

我们研究了为包含数百个物体的杂乱场景生成组合式3D表征的问题。其目标是将场景表示为放置在统一世界坐标系下的独立物体网格集合，以满足游戏、AR/VR、仿真和机器人等下游应用的需求。在密集杂乱的场景中，该任务极具挑战性，因为物体之间存在严重的相互遮挡，且每个视角仅能展现其几何结构的一小部分。基于几何的方法通常将场景重建为单一表征，并在遮挡区域留下不完整的几何形状；而现有的结合生成先验的组合式方法则大多局限于相对简单的场景。我们表明，通过将强大的单物体3D生成先验适配到多视角观测中，完全可以采用组合式方法生成包含数百个物体的复杂场景。我们通过 Pixal3D 实现了这一范式，为其扩展了一条多视角条件路径，将物体生成锚定在多个带位姿的观测图像中。尽管该模型完全是在规范空间中的单个物体上进行微调的，但在没有任何场景级训练的情况下，它依然能够泛化到存在严重遮挡的大规模场景中，这证明了该范式的可行性与可扩展性。我们进一步推出了 UE-MeshyScene，这是一个照片级逼真的密集杂乱场景基准，包含数百个物体，并提供逐物体的标注及真值网格。在单物体、受控多物体以及 UE-MeshyScene 的各项评估中，我们的方法均持续优于以往方法，且随着场景复杂度和遮挡程度的增加，优势更为显著。最后，我们通过将生成的 3DGS 世界（如 Marble 和 HY-World 2.0）转换为组合式网格场景，展示了该方法更广泛的适用性。