Reproduce “DeepEMD: A Transformer-based Fast Estimation of the Earth Mover's Distance” 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
Atul Kumar Sinha, Francois Fleuret
Atul Kumar Sinha and François Fleuret from the University of Geneva introduce DeepEMD, a transformer-based neural network that accurately estimates the Earth Mover's Distance (EMD) between point clouds by predicting the underlying bipartite matching. This approach transforms EMD from a computationally prohibitive metric into a practical training loss for deep generative models, enabling improved point cloud reconstruction quality.
Key insights
- Learning to predict the optimal bipartite matching directly allows for both accurate EMD estimation and reliable gradient computation, making it suitable as a training loss.
- Transformer architectures with attention mechanisms are effective for capturing the complex global dependencies required to solve the optimal transport problem between point clouds.
- Accurate approximation of EMD can enhance the quality of 3D point cloud generative models, leading to better geometric and perceptual reconstructions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.