🎨 Homepage

NeuralPlane

Mon, 24 Feb 2025 15:20:29 GMT

Hanqiao Ye1,2, Yuzhou Liu1,2, Yangdong Liu2, Shuhan Shen1,2
School of Artificial Intelligence, University of Chinese Academy of Sciences1 , Institute of Automation, Chinese Academy of Sciences2
Paper Supplementary Code (To be uploaded) NeuralPlane rebuilds indoor scenes as arrangements of planar primitives.
3D maps assembled from planar primitives are compact and expressive in representing man-made environments, making them suitable for a spectrum of applications. In this paper, we present NeuralPlane, a novel approach that explores neural fields for multi-view 3D plane reconstruction. Our method is centered upon the core idea of distilling geometric and semantic cues from inconsistent 2D plane observations into a unified 3D neural representation, which unlocks the full leverage of plane attributes. NeuralPlane realizes this idea via several key designs, including: 1) a monocular module that generates geometrically smooth and semantically meaningful segments as 2D plane observations, 2) a plane-guided training procedure that implicitly learns accurate plane locations from multi-view plane observations, and 3) a self-supervised feature field termed Neural Coplanarity Field that enables the modeling of scene semantics alongside the geometry. Without relying on plane annotations, our method achieves high-fidelity reconstruction comprising planar primitives that are not only crisp but also well-aligned with the semantic content. Comprehensive experiments on ScanNetv2 and ScanNet++ demonstrate the superiority of our results in both geometry and semantics ReferenceChoose a result from PlanarRecon ObjectSDF++AirPlanesOursReferenceChoose a result from PlanarRecon ObjectSDF++AirPlanesOursReferenceChoose a result from PlanarRecon ObjectSDF++AirPlanesOursReferenceChoose a result from PlanarRecon ObjectSDF++AirPlanesOurs@inproceedings{ ye2025neuralplane, title={NeuralPlane: Structured 3D Reconstruction in Planar Primitives with Neural Fields}, author={Hanqiao Ye and Yuzhou Liu and Yangdong Liu and Shuhan Shen}, booktitle={The Thirteenth International Conference on Learning Representations}, year={2025}, url={https://openreview.net/forum?id=5UKrnKuspb} }

video_compare

Mon, 24 Feb 2025 14:51:06 GMT

results_gallery

Mon, 24 Feb 2025 14:46:31 GMT

NeuralPlane

Mon, 24 Feb 2025 12:46:52 GMT

Hanqiao Ye1,2, Yuzhou Liu1,2, Yangdong Liu2, Shuhan Shen1,2
School of Artificial Intelligence, University of Chinese Academy of Sciences1 , Institute of Automation, Chinese Academy of Sciences2
Paper Supplementary Code (To be uploaded) NeuralPlane rebuilds indoor scenes as arrangements of planar primitives.
3D maps assembled from planar primitives are compact and expressive in representing man-made environments, making them suitable for a spectrum of applications. In this paper, we present NeuralPlane, a novel approach that explores neural fields for multi-view 3D plane reconstruction. Our method is centered upon the core idea of distilling geometric and semantic cues from inconsistent 2D plane observations into a unified 3D neural representation, which unlocks the full leverage of plane attributes. NeuralPlane realizes this idea via several key designs, including: 1) a monocular module that generates geometrically smooth and semantically meaningful segments as 2D plane observations, 2) a plane-guided training procedure that implicitly learns accurate plane locations from multi-view plane observations, and 3) a self-supervised feature field termed Neural Coplanarity Field that enables the modeling of scene semantics alongside the geometry. Without relying on plane annotations, our method achieves high-fidelity reconstruction comprising planar primitives that are not only crisp but also well-aligned with the semantic content. Comprehensive experiments on ScanNetv2 and ScanNet++ demonstrate the superiority of our results in both geometry and semantics ReferenceChoose a result from OursReferenceChoose a result from OursReferenceChoose a result from OursReferenceChoose a result from Ours@inproceedings{ ye2025neuralplane, title={NeuralPlane: Structured 3D Reconstruction in Planar Primitives with Neural Fields}, author={Hanqiao Ye and Yuzhou Liu and Yangdong Liu and Shuhan Shen}, booktitle={The Thirteenth International Conference on Learning Representations}, year={2025}, url={https://openreview.net/forum?id=5UKrnKuspb} }

teaser

Tue, 01 Oct 2024 07:30:20 GMT

Home

Sat, 22 Feb 2025 07:59:45 GMT

Hi there 👋 , this is YE HANQIAO (/yeh han chow/, 叶翰樵 ), a third-year PhD student at School of Artificial Intelligence, University of Chinese Academy of Sciences (UCAS). I am conducting reseach at 3DV Lab of Institute of Automation, Chinese Academy of Sciences (CASIA) supervised by Professor Shuhan Shen. Prior to that, I received my B.E. degree from Harbin Institute of Technology, Shenzhen (HITSZ) in 2022, advised by Professor Jun Xu. My research interest lies in the 3D Computer Vision, particularly including 3D representation, pose estimation and scene understanding, as well as their applications in mixed reality and robotics. Currently, I am deeply involved in specific projects related to visual localization. If you are interested in my work, please feel free to drop me an e-mail.
[2025/02] One paper is to be presented as an Oral in ICLR 2025 🥳. [2024/07] A co-authored paper gets accepted in ECCV 2024.
[2024/06] We win first place in the 1st Building3D challenge in CVPR 2024 Workshop on Urban Scene Modeling. [2024/04] A co-authored paper submitted to IEEE RA-L gets accepted. WarningThis webpage is VERY much a work in progress. As such it is incomplete.

home-banner

Sat, 22 Feb 2025 07:42:19 GMT

Portrait

Fri, 21 Feb 2025 08:50:13 GMT

teaser-video

Sat, 28 Sep 2024 05:29:01 GMT

bulma.css.map

Tue, 24 Sep 2024 01:42:53 GMT

index

Mon, 24 Feb 2025 08:49:02 GMT