具有透視感知的室內家具物件層級影像客製化

Abstract

本研究提出一套具備透視感知能力的室內家具影像客製化方法，旨在提升物件插入結果的幾何一致性與視覺真實感。我們擴展 AnyDoor 框架，加入「Perspective BBox Branch」，以消失點對齊的 3D 邊界框投影作為條件輸入，引導 diffusion 模型生成符合場景透視的合成結果。同時，我們設計了互動式使用者介面，使用戶能標註消失點、生成一致的 3D BBox，並繪製參考物件遮罩，實現可控且直覺的物件放置流程。 我們使用 ScanNet++ 室內資料集進行模型訓練與驗證，實驗結果顯示，本方法在不犧牲視覺品質的情況下，有效提升透視一致性。量化指標（CLIP 與 DINO 分數）與質化視覺結果皆支持本方法的有效性。此外，透過對於未見資料（out-of-distribution）場景的測試，我們證實模型具備良好泛化能力。本研究展示了幾何感知引導在生成式模型中的潛力，為未來具備控制性與結構意識的影像合成方法奠定基礎。

We propose a perspective-aware framework for object-level image customization in indoor scenes, aiming to improve geometric consistency and visual realism in object insertion tasks. Building upon the AnyDoor architecture, we introduce the Perspective BBox Branch, which encodes vanishing-point-aligned 3D bounding box projections as conditional guidance for the diffusion model. To support controllable and user-friendly annotation, we also develop an interactive user interface that allows users to mark vanishing points, generate perspective-aligned 3D BBoxes, and draw reference object masks. We train and evaluate our model on the ScanNet++ dataset, and experimental results demonstrate that our method significantly improves perspective alignment without compromising visual fidelity. Quantitative metrics (CLIP and DINO scores) and qualitative visualizations both confirm the effectiveness of our approach. Furthermore, tests on out-of-distribution indoor images show that the model generalizes well to real-world scenarios. Our work highlights the potential of geometry-aware conditioning in generative models and offers a practical foundation for future research on controllable indoor image synthesis.