Jiachen Liu

I'm currently a machine learning engineer at Tiktok in San Jose. I received my PhD in Information Sciences and Technology from Penn State University in 2024, where I am fortunately to be advised by Dr. Sharon X. Huang. Prior to that, I did my bachelor degree in Beihang University in China. My research interests and experience include 3D vision and generative AI. Specifically, I am interested in structured 3D reconstruction, scene understanding and their downstream applications, as well as generalizable 3D recontruction across various environments.

Email / Google Scholar / Linkedin / Github

Selected Research

My primary interest lies in 3D vision and generative AI, including structured 3D reconstruction, scene understanding, generalizable 3D vision, scene layout generation and 3D-aware asset generation. (* indicates equal contribution.)

	Towards In-the-wild 3D Plane Reconstruction from a Single Image Jiachen Liu, Rui Yu, Sili Chen, Sharon X. Huang, Hengkai Guo CVPR, 2025 (Highlight Presentation) code / arXiv We aim to propose the problem of in-the-wild, zero-shot 3D plane reconstruction. To this end, (1) We have constructed a large-scale benchmark dataset with high-quality dense planar annotations from multiple RGB-D datasets sampled across various indoor and outdoor environments. (2) We propose a Transformer-based framework on mixed-dataset training with a disentangled, classification-then-regression normal and offset learning paradigm to effectively handle the challenge raised in scale invariance across diverse indoor and outdoor scenes. Our model has demonstrated state-of-the-art planar reconstruction performance in terms of both accuracy and generalizability.
	Computer-Aided Layout Generation for Building Design: A Review Jiachen Liu, Yuan Xue, Haomiao Ni, Rui Yu, Zihan Zhou, Sharon X. Huang; CVMJ, 2025 arXiv We present a comprehensive survey on current methods on floorplan layout generation, scene synthesis and other miscellaneous layout generation topics. We also discuss valuable future perspectives to work on in this area.
	MonoPlane: Exploiting Monocular Geometric Cues for Generalizable 3D Plane Reconstruction Wang Zhao, Jiachen Liu, Sheng Zhang, Yishu Li, Sili Chen, Sharon X. Huang, Yong-Jin Liu, Hengkai Guo; IROS, 2024 (Oral Presentation) code (Coming soon) / arXiv Leverage pretrained geometric foundation models (depth and normal) for zero-shot monocular plane reconstruction.
	NeRF-Enhanced Outpainting for Faithful Field-of-View Extrapolation Rui Yu, Jiachen Liu, Zihan Zhou, Sharon X. Huang ICRA, 2024 arXiv We use NeRF to represent an indoor scene for novel view synthesis to augment the training set, then train a view extrapolation network on the densely sampled synthesized data to improve the model's view extrapolation capacity in a 3D coherent manner.
	End-to-end Graph-constrained Vectorized Floorplan Generation with Panoptic Refinement Jiachen Liu, Yuan Xue, Jose Duarte, Krishnendra Shekhawat, Zihan Zhou, Xiaolei Huang ECCV*, 2022 arXiv We propose a Transformer-based two-stage framework to effecitvely generate vectorized interior floorplan layouts in an end-to-end manner.
	PlaneMVS: 3D Plane Reconstruction from Multi-View Stereo Jiachen Liu, Pan Ji, Nitin Bansal, Changjiang Cai, Qingan Yan, Xiaolei Huang, Yi Xu CVPR, 2022 code / arXiv We present a multi-view plane reconstruction framework based on the proposed slanted plane hypothesis, where we achieve state-of-the-art plane reconstruction as well as superior multi-view depth estimation across multiple benchmark datasets.

Academic Service

I serve as a regular reviewer for top-tier CV/ML conferences, such as CVPR, ICCV, ECCV, NeurIPS, ICLR, ICML, etc.

Design and source code from Jon Barron's website