Reproduce “Mover360: Controllable Object Manipulation in 360° Panoramic Images” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Haoyi Zhong, Fang-Lue Zhang, Andrew Chalmers, Taehyun Rhee
Mover360, developed by researchers from Victoria University of Wellington, UNSW, and The University of Melbourne, provides a controllable framework for object manipulation in 360-degree equirectangular panoramic images, achieving superior quantitative performance and generating globally consistent edits for object translation, removal, and insertion across synthetic and real-world datasets.
Key insights
- Direct manipulation within the ERP domain, combined with geometric conditioning (depth maps) and architectural adaptations (circular padding), is essential for preserving global consistency and avoiding artifacts in 360-degree panoramic edits.
- A large, diverse paired dataset, generated through a sophisticated 3D pipeline like UE5, is critical for training robust models capable of complex object manipulation tasks in panoramic environments.
- A unified, LoRA-adapted pretrained diffusion transformer can effectively handle multiple object manipulation tasks with varying levels of user guidance (point, bounding box, mask) within a single model.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.