Reproduce “RCR: Robust Crowd Reconstruction with Upright Space from a Single Large-scene Image” 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
Jing Huang, Hao Wen, Tianyi Zhou, Haozhe Lin, Yu-kun Lai, Kun Li
The RCR method presents a top-down pipeline for robustly reconstructing 3D poses, shapes, and absolute locations of hundreds of people from a single large-scene image, even with unknown and arbitrary camera fields of view. This approach significantly outperforms prior methods in spatial consistency and pose/shape accuracy, and introduces two large-scale datasets for crowd reconstruction.
Key insights
- The Human-scene Virtual Interaction Point (HVIP) effectively transforms the complex 3D human localization problem into a more manageable 2D estimation task.
- Creating canonical 'Upright 3D/2D Spaces' and applying 'Upright Normalization' to inputs effectively removes the influence of camera parameters and cropping, enhancing the generalization of single-person reconstruction models.
- An iterative, ground-aware cropping strategy can dynamically adjust to varying human scales and unknown camera/ground parameters, resolving the circular dependency in large-scene human detection and improving overall robustness.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.