Depth estimation method based on CF-ACV stereo matching network
Year: 2025
Type: article
Abstract: Deep learning-based binocular stereo matching methods, although capable of maintaining high accuracy, often struggle to balance computational efficiency and exhibit limited cross-domain generalization, thereby hindering their direct application in new real-world scenarios. Existing stereo matching networks tend to lose a portion of geometric and contextual information during the feature channel generation process. The effective utilization of such information not only enhances the accuracy of matching results but also improves the efficiency of feature extraction and the overall matching process. Moreover, the comprehensive integration of geometric and contextual cues can reduce the model’s sensitivity to scale and viewpoint variations, thereby significantly improving its generalization capability. To address these challenges, a stereo matching network named CF-ACV has been proposed, which fuses contextual and geometric information to optimize the cost aggregation strategy. This integration enables the model to achieve high matching accuracy while maintaining superior computational efficiency and exceptional generalization performance. Building upon CF-ACV, a high-efficiency, high-precision, and all-scenario depth estimation method has been developed. Experimental results indicate that, compared to traditional stereo matching methods, the proposed approach achieves an average improvement in matching accuracy of approximately 43.19\%, and outperforms most real-time state-of-the-art stereo matching networks in both speed and precision. In addition, the manual marking of feature points in highly reflective regions further refines the model's performance in weak-texture and high-reflectivity areas, resulting in smoother and more continuous disparity maps. Finally, experimental evaluations of the constructed depth estimation method across three different scenarios yielded an average absolute error of 6.44 mm and an average relative error of 4.15\%, thereby demonstrating that the proposed method is capable of efficient and accurate depth estimation.
Source: Optics and Lasers in Engineering
Authors: Xurong Wang, Wenxin Hu
Institution: Shenzhen MSU-BIT University
Language: English
Topic: Advanced Vision and Imaging
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $4,020
Awards: G8682481250, G8957477452
OpenAlex ID: W4411275195