Reproduce “LUVLi Face Alignment: Estimating Landmarks' Location, Uncertainty, and Visibility Likelihood” 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
Abhinav Kumar, Tim K. Marks, Wenxuan Mou, Ye Wang, Michael Jones, Anoop Cherian, Toshiaki Koike-Akino, Xiaoming Liu and 1 more
LUVLi estimates facial landmarks while predicting location, spread, and visibility. It derives subpixel locations from heatmaps and trains location, uncertainty, and visibility predictions. On COFW-68, its uncertainty is 32.49 for externally occluded landmarks versus 9.31 for unoccluded points, while localization errors are lower than KDN’s in both groups. This shows separation, not calibrated uncertainty or guaranteed failure detection.
Key insights
- Parametric uncertainty modeling, particularly using a Laplacian distribution for 2D landmark locations, provides highly predictive measures of localization error.
- Jointly estimating visibility with uncertainty and location enhances robustness, allowing the model to differentiate between various occlusion types and manage self-occluded landmarks effectively.
- Utilizing a differentiable weighted spatial mean of ReLU-activated heatmap elements enables sub-pixel landmark localization accuracy within an end-to-end trainable framework.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.