Reproduce “Sparse Beats Dense: Rethinking Supervision in Radar-Camera Depth Completion” 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
Huadong Li, Minhao Jing, Jiajun Liang, Haoqiang Fan, Renhe Ji
A radar-camera depth completion framework shows that clean, single-frame sparse LiDAR supervision can outperform pseudo-dense labels when sensor-pattern shortcuts are controlled. By disrupting LiDAR distribution leakage and restoring useful radar geometry and semantic cues, the method reaches 1,927.0 mm MAE on nuScenes at 0–80 m, an approximately 11.6% improvement over RadarNet, while running at 21.7 FPS.
Key insights
- More supervised pixels are not necessarily better: noisy pseudo-dense labels can be less useful than accurate single-frame sparse labels when the model’s shortcut behavior is addressed.
- Camera and radar disruptions work best together; vertical radar-line extension is more effective than adding noise independently to radar height or depth coordinates.
- Object-level MAE exposes stripe artifacts that conventional sparse-pixel MAE and RMSE may miss, although it remains an indirect measure because reference depths are still sparse.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.