A CVPR Oral stereo matcher built on MegEngine

What it does CREStereo is the official MegEngine implementation of a CVPR 2022 Oral paper for stereo matching. It takes left and right stereo images and estimates a disparity map through a cascaded recurrent network. The repository also releases the paper’s custom ~400GB training dataset, alongside support for standard benchmarks like KITTI, Middlebury, and SceneFlow.

The interesting bit Rather than a single-shot estimator, the architecture repeatedly refines its disparity guess in stages—code lineage traces back to RAFT’s recurrent backbone and LoFTR’s attention module. The README doesn’t unpack the “adaptive correlation” trick, so you’ll need the paper for the gory details.

Key highlights

• Official implementation tied to MegEngine v1.8.2, not PyTorch or TensorFlow
• Pretrained MegEngine model and a Docker image for quick inference
• Releases a ~400GB custom training dataset plus loaders for public stereo datasets
• Training and test scripts included with TensorBoard logging
• Code base borrows from RAFT, LoFTR, and HSMNet

Caveats

• Dependencies are frozen in time: CUDA 10.1, Python 3.6.9, and MegEngine v1.8.2
• The README lists no quantitative benchmarks or comparisons against other methods
• “Adaptive correlation” and the cascaded design are named but not explained; the paper is required reading

Verdict Grab it if you need a reproducible CVPR Oral baseline in the MegEngine ecosystem or want the custom dataset. Skip it if you were hoping for a modern PyTorch port or a plug-and-play production tool.