📝 Publications

CompoNeRF: Text-guided Multi-object Compositional NeRF with Editable 3D Scene Layout

Haotian Bai, Yuanhuiyi Lyu, Lutao Jiang, Sijia Li, Haonan Lu, Xiaodong Lin, Lin Wang

[Project] | [Video] |

• A novel framework that synthesizes coherent multi-object scenes by integrating textual descriptions with box-based spatial arrangements.
• CompoNeRF is designed for precision and adaptability, allowing for individual NeRFs, each denoted by a unique prompt color, to be composed, decomposed, and recomposed with ease, streamlining the construction of complex scenes from cached models after decomposition.

Dynamic PlenOctree for Adaptive Sampling Refinement in Explicit NeRF

Haotian Bai, Yiqi Lin, Yize Chen, Lin Wang

Project | Video |

• A more compact and fertile PlenOctree (POT) NeRF representation.
• Inspiration: POT’s fixed structure for direct optimization is sub-optimal as the scene complexity evolves continuously with updates to cached color and density, necessitating refining the sampling distribution to capture signal complexity accordingly.
• Competitive: DOT outperforms POT by enhancing visual quality, reducing over 55.15/68.84% parameters, and providing 1.7/1.9 times FPS for NeRF-synthetic and Tanks and Temples.

Patch-Mix Transformer for Unsupervised Domain Adaptation: A Game Perspective

Jinjing Zhu*, Haotian Bai^*, Lin Wang

Project | Video |

• Be selected as one of CVPR (highlight) papers(top 2.5%)
• Large Domain Gap: PMTrans bridges source and target domains with an intermediate domain in a relatively smooth way.
• Game Theory: Interpret UDA as a min-max CE game with three players, including the feature extractor, classifier, and PatchMix to find the Nash Equilibria.
• Competitive: PMTrans surpasses ViT-based and CNN-based SoTA methods by +3.6% on Office-Home, +1.4% on Office-31, and +17.7% on DomainNet.

Weakly Supervised Object Localization via Transformer with Implicit Spatial Calibration

Haotian Bai, Ruimao Zhang, Jiong Wang, Xiang Wan

Project | Video |

• SCM is the external transformer based solution for Weakly Supervised Object Localization.
• Lightweight: SCM is an external Transformer model that produces no additional parameters.
• Competitive: SCM outperforms most competitive frameworks (CNN & Transformer) using only about 𝟐𝟎%~𝟑𝟎% of their parameters.