publications

2026

1. ACM MM

   VC-Agent: An Interactive Agent for Customized Video Dataset Collection

   Yidan Zhang, Mutian Xu, Yiming Hao, and 6 more authors

   In ACM International Conference on Multimedia, 2026

   Abs DOI arXiv Website

   Customized video datasets are difficult to scale because users rarely know how to state every requirement in a single query. VC-Agent is the first interactive MLLM agent that turns lightweight feedback into an evolving collection policy, using template-based acceptance and attribute-aware rejection. Its end-to-end system combines a Web UI, video proposal, iterative criteria refinement, filtering, and automatic scale-up. In user studies, participants collected 335K videos across eight datasets with only minutes of interaction per dataset and 84–92% judged quality.

2. ICLR

   ReconViaGen: Towards Accurate Multi-view 3D Object Reconstruction via Generation

   Jiahao Chang, Chongjie Ye, Yushuang Wu, and 6 more authors

   In International Conference on Learning Representations, 2026

   Abs arXiv Website

   Pure reconstruction is accurate where views overlap but incomplete elsewhere, while diffusion models can complete unseen regions yet drift away from the input images. ReconViaGen conditions a 3D diffusion generator on global and per-view reconstruction priors, then uses rendering-aware velocity compensation to preserve geometric and textural alignment. It achieves state-of-the-art results on Dora-Bench and OmniObject3D.

2023

1. CVPR

   MVImgNet: A Large-scale Dataset of Multi-view Images

   Xianggang Yu, Mutian Xu, Yidan Zhang, and 10 more authors

   In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023

   Abs Website

   Large-scale real-world 3D data is expensive to collect, leaving 3D learning without an ImageNet-scale source of multi-view observations. MVImgNet contains 6.5M frames from 219,188 videos across 238 classes, with object masks, camera parameters, and point clouds. It supports radiance-field reconstruction, multi-view stereo, and view-consistent understanding, and further yields MVPNet, a real-world point-cloud dataset with 87,200 objects.