Reproduce “HITNet: Hierarchical Iterative Tile Refinement Network for Real-time Stereo Matching” 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
Vladimir Tankovich, Christian Häne, Yinda Zhang, Adarsh Kowdle, Sean Fanello, Sofien Bouaziz
HITNet, developed by Google, presents a neural network architecture for real-time stereo matching that sidesteps full 3D cost volumes, instead employing hierarchical iterative tile refinement. It achieves competitive accuracy with runtimes as low as 19ms on KITTI benchmarks and strong generalization to real-world datasets.
Key insights
- Integrating explicit geometric reasoning, like slanted plane hypotheses and spatial propagation from classical stereo, into a neural network can achieve high accuracy while avoiding the computational burden of 3D cost volumes.
- Hierarchical processing with multi-resolution initialization and iterative refinement allows the network to maintain high resolution and reconstruct fine details efficiently.
- Learning compact tile feature descriptors alongside geometric parameters enables a flexible and data-driven approach to local surface approximation and matching.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.