Reproduce “Sparse and noisy LiDAR completion with RGB guidance and uncertainty” 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
Wouter Van Gansbeke, Davy Neven, Bert De Brabandere, Luc Van Gool
FusionNet leverages a dual-branch neural network and explicit uncertainty estimation to complete sparse and noisy LiDAR data into dense depth maps with RGB guidance. It achieved an RMSE of 773 mm on the KITTI depth completion benchmark and operates at 50 frames per second.
Key insights
- Combining global contextual information (derived from RGB and LiDAR) with local fine-grained processing through both early and uncertainty-weighted late fusion significantly enhances depth completion accuracy.
- Explicitly learning and leveraging confidence maps enables the network to adaptively prioritize the more reliable depth estimations from either the global or local branch at a pixel level.
- Strategic omission of batch normalization in the initial layers of the local network helps mitigate issues caused by extreme input sparsity, preventing parameter skewing.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.