Title: TASE: Truncation-Aware Semantic Embeddings for 3D Scene Understanding and Editing

Abstract:

High-quality semantic 3D scene representations are essential for a wide range of applications, such as simulation, autonomous driving, and robotics. Furthermore, the capacity to modify these representations allows developers to tailor applications to specific scenarios with greater ease. However, existing methods offer only constrained support for controllable editing. To address this, we present TASE, a novel approach that maps pretrained 2D semantic features into a truncation-aware embedding space, thereby facilitating flexible editing of 3D scenes.

Our technique explicitly optimizes the feature space so that progressively reducing the number of feature channels generates increasingly abstract semantic representations, whereas retaining a larger number of channels preserves fine-grained details. We also enhance the multi-view consistency of these features by employing a loss function based on scale and translation equivariance. This resulting embedding space supports text-driven edits to 3D scenes, offering explicit control over the degree to which edits align with the original scene content. Consequently, our method permits more significant modifications than previous approaches. Additionally, we introduce a finetuning stage for the editing diffusion model to reduce artifacts arising from geometric changes. Experimental outcomes show competitive performance in 3D scene editing, with our method substantially surpassing prior techniques in scenarios involving large geometric modifications.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC