Reproduce “A New People-Object Interaction Dataset and NVS Benchmarks” on OpenResearch

About this paper

Shuai Guo, Houqiang Zhong, Qiuwen Wang, Ziyu Chen, Yijie Gao, Jiajing Yuan, Chenyu Zhang, Rong Xie and 1 more

Researchers from Shanghai Jiao Tong University introduced a comprehensive multi-person, multi-view RGB-D human-object interaction dataset captured with 30 synchronized 4K cameras, augmented with extensive prior information. Benchmarks on this new dataset demonstrate that current state-of-the-art novel view synthesis models exhibit significant overfitting when applied to unseen dynamic interaction views.

Key insights

• A high-fidelity, multi-modal, multi-view dataset featuring dynamic interactions is essential for advancing research in human-object interaction understanding and novel view synthesis.
• Current state-of-the-art novel view synthesis models, such as TensoRF, K-Planes, and 3D Gaussian Splatting, demonstrate significant overfitting when tested on unseen views from complex, dynamic interaction scenes.
• The provision of rich prior information, including SMPL models, point clouds, and meshes, is critical for addressing reconstruction challenges and mitigating overfitting in NVS models dealing with occlusions and complex poses.

Summary by alphaXiv.