AssetField - Assets Mining and Reconfiguration in Ground Feature Plane Representation

Aug 23, 2026

Upload by Bintang

AssetField addresses the need for flexible, user-friendly editing of structured indoor and outdoor scenes in virtual environments. The method learns object-aware ground feature planes in a bird-eye view aligned 2D representation, enabling natural visualization and intuitive manipulation such as translation, duplication, rotation, deformation, and deletion. An unsupervised asset library of template feature patches supports category-aware grouping edits and reduces repetitive work. Results demonstrate strong novel-view synthesis and realistic renderings for new scene configurations.

LanguageEnglish

CategoryTechnology

FormatPDF

Pages10

Reading Time25 min

AccessFree to download

Introduction

Problem and motivation

Limitations of existing neural rendering approaches

Proposed representation and contributions

AssetField: Assets Mining and Reconﬁguration in Ground Feature Plane Representation Yuanbo Xiangli 1 􀀃 , Linning Xu 1 􀀃 , Xingang Pan3 , Nanxuan Zhao4 , Bo Dai2􀀌, Dahua Lin 1 ;2 1 The Chinese University of Hong Kong 2 Shanghai AI Laboratory 3 Max Planck Institute for Informatics 4 Adobe Research fxy019,xl020,[dhlin](dhling@ie.cuhk.edu.hk)[g](dhling@ie.cuhk.edu.hk)[@ie.cuhk.edu.hk](dhling@ie.cuhk.edu.hk) [xpan@mpi-inf.mpg.de](xpan@mpi-inf.mpg.de)[nanxuanzhao@gmail.com](nanxuanzhao@gmail.com) [daibo@pjlab.org.cn](daibo@pjlab.org.cn) Figure 1: Man-made environments are often characterized by repetitive scene objects, e.g. tables, chairs, and trees. AssetField represents these environments with a set of informative ground feature planes aligning with the physical ground, from which neural representations of scene objects are extracted and grouped into categories. The proposed mechanism allows users to manipulate and compose assets directly on the ground feature plane and produces high-quality rendering on novel scene conﬁgurations. Abstract Both indoor and outdoor environments are inherently structured and repetitive. Traditional modeling pipelines keep an asset library storing unique object templates, which is both versatile and memory efﬁcient in practice. Inspired by this observation, we propose AssetField, a novel neural scene representation that learns a set of object-aware ground feature planes to represent the scene, where an asset library storing template feature patches can be constructed in an unsupervised manner. Unlike existing methods which require object masks to query spatial points for object editing, our ground feature plane representation offers a natural visualization of the scene in the bird-eye view, allowing a variety of operations (e.g. translation, duplication, deformation) on objects to conﬁgure a new scene. With the template feature patches, group editing is enabled for scenes with many recurring items to avoid repetitive work on object individuals. We show that AssetField not only achieves competitive performance for novel-view synthesis but also generates realistic renderings for new scene conﬁgurations. 1. Introduction The demand for bringing our living environment into a virtual realm continuous to increase these days, with example cases ranging from indoor scenes such as rooms and restaurants, to outdoor ones like streets and neighborhoods. Apart from the realistic 3D rendering, real-world applications also require ﬂexible and user-friendly editing of the scene. Use cases can be commonly found in interior design, urban planning etc. To save human labor and expense, users need to frequently visualize different scene conﬁgurations before ﬁnalizing a plan and bringing it to reality, like shown in Fig.1. For their interests, a virtual environment offering versatile editing choices and high rendering quality is always preferable. In these scenarios, objects are primarily located on a horizontal plane like ground, and can be inserted to/deleted from the scene. Translation along the plane and rotation around the vertical axis are also common operations. Furthermore, group editing becomes essential when scenes are populated with recurring items (e.g. substitute all chairs with stools and remove all sofas in a bar) . While recent advances in neural rendering [27, 3, 44, 28] offer promising solutions to producing realistic visuals, they struggle to meet the aforementioned editing demands. Speciﬁcally, traditional neural radiance ﬁeld (NeRF)-based methods such as [47, 26, 3] encode an entire scene into a single neural network, making it difﬁcult to manipulate and composite due to its implicit nature and limited network capacity. Some follow-up works [41, 16] tackle objectaware scene rendering in a bottom-up fashion by learning one model per object and then performing joint rendering. Another branch of methods learn object radiance ﬁelds using instance masks [42], object motions [46], and ima

FAQs

What problem does AssetField aim to solve in virtual scene creation?

AssetField targets the gap between realistic neural rendering and practical editing needs, including flexible, user-friendly operations on scene objects.

How does AssetField represent scenes to support editing?

It factorizes a 3D neural field into a bird-eye-view-aligned ground feature plane plus a vertical feature axis, enabling intuitive object-level manipulation.

How are group editing operations supported?

AssetField maintains a cross-scene asset library of template feature patches, allowing category-level grouping edits for scenes with many recurring items.

Aug 23, 2026

Upload by Bintang