Match Stereo Videos via Bidirectional Alignment

Area of Science:

• Computer Vision
• Machine Learning
• 3D Reconstruction

Background:

• Video stereo matching aims to create consistent disparity maps from stereo videos.
• Current methods often lack temporal consistency and struggle with specific environments.
• Existing techniques may produce artifacts like low-frequency oscillations.

Purpose of the Study:

• To develop advanced methods for video stereo matching with improved temporal consistency.
• To introduce new datasets for training and evaluating stereo matching in natural and urban scenes.
• To enhance the performance and robustness of disparity map estimation.

Main Methods:

• A bidirectional alignment mechanism for adjacent frames.
• A novel video processing framework called BiDAStereo.
• A plugin stabilizer network, BiDAStabilizer, for general image-based methods.
• Creation of realistic synthetic and real-world datasets for natural and urban scenes.

Main Results:

• Demonstrated improvements in prediction quality and temporal consistency.
• Achieved state-of-the-art results on multiple benchmarks.
• Validated the effectiveness of BiDAStereo and BiDAStabilizer across various evaluations.
• Showcased the utility of the new datasets for robust benchmarking.

Conclusions:

• The proposed bidirectional alignment and BiDAStereo framework significantly advance video stereo matching.
• The new datasets fill a critical gap, enabling better evaluation in diverse, real-world scenarios.
• The methods offer improved disparity map accuracy and temporal stability, outperforming existing approaches.