Reproduce “Adaptive Fusion of Single-View and Multi-View Depth for Autonomous Driving” 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
JunDa Cheng, Wei Yin, Kaixuan Wang, Xiaozhi Chen, Shijie Wang, Xin Yang
Researchers from Huazhong University of Science and Technology and DJI Technology developed AFNet, a depth estimation system that adaptively fuses single-view and multi-view cues to maintain high accuracy and superior robustness, particularly when camera poses are noisy. The method achieved state-of-the-art performance on DDAD and KITTI datasets, reducing AbsRel error on KITTI by 27.8% compared to previous methods, and demonstrated improved precision in dynamic object regions.
Key insights
- Robust depth estimation requires explicit handling of noisy camera poses, which is a critical, often overlooked, challenge in real-world autonomous driving.
- Adaptive, pixel-wise fusion, guided by confidence measures including a geometrically-derived warping confidence, allows a system to dynamically leverage the strengths of both single-view and multi-view cues based on their perceived reliability.
- The development of dedicated robustness benchmarks (synthetic and real-world SLAM noise) is essential for assessing the practical viability and improving the reliability of depth estimation algorithms for deployment.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.