CREStereo: Stereo Image Pairs and Disparity Maps for Depth Estimation

CREStereo provides stereo image pairs and ground truth disparity maps developed by Megvii Research for high-resolution stereo matching tasks. Originally released in 2022, the data supports the training of cascaded recurrent networks to handle complex geometric structures and reflections.

Use Cases

• Training neural networks for disparity estimation using left and right image inputs
• Benchmarking depth estimation accuracy against ground truth disparity maps
• Developing algorithms to handle reflections and thin structures in stereo vision

Strengths

• Includes ground truth disparity maps
• Developed for CVPR 2022 research
• Focuses on high-resolution stereo matching

Limitations

• Metadata does not specify the exact number of image pairs
• Potential domain shift between synthetic training samples and real-world applications

Provenance

Source

Megvii Research

Collection Method

Synthetic generation and real-world capture

Time Range

2022

Freshness

Based on the 2022 CVPR publication by Megvii Research.