Reproduce “UniLiPs: Unified LiDAR Pseudo-Labeling with Geometry-Grounded Dynamic Scene Decomposition” 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
Filippo Ghilotti, Samuel Brucker, Nahku Saidy, Matteo Matteucci, Mario Bijelic, Felix Heide
The UniLiPs framework introduces an unsupervised method for concurrently generating high-quality pseudo-labels for 3D semantic segmentation, object detection, and depth estimation from unlabeled LiDAR data. By leveraging geometry-grounded dynamic scene decomposition, the approach improves depth prediction accuracy at long ranges and enables state-of-the-art perception models to achieve near-oracle performance on various tasks, addressing the challenge of expensive 3D data annotation for autonomous driving.
Key insights
- A unified approach to generating pseudo-labels for multiple 3D perception tasks simultaneously, built upon a consistent SLAM map, can effectively overcome the limitations of task-specific annotation methods.
- Combining image-based foundational models with geometry-aware probabilistic constraints allows for robust semantic labeling and dynamic object identification in complex, real-world driving scenarios.
- Iterative refinement, including the removal of 'floaters' (dynamic objects mistakenly registered as static) and label propagation, is crucial for improving the quality, accuracy, and density of generated pseudo-labels.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.