A Novel Stereo Vision Universality Algorithm Model Suitable for Tri-PSMNet and InvTri-PSMNet

This study investigates the feasibility of sharing weight parameters between L-shaped and inverted triangle three-camera configurations. An optimized universal framework was developed. Initially, Tri-PSMNet using InvTri-PSMNet trained weights achieved an error rate of 40.2% on the Tri-Scene test, while InvTri-PSMNet using Tri- Psmn et trained weights achieved an error rate of 37.1% on the InvTri-Scene test. After optimization, the universal framework, without requiring weight training, directly utilized InvTri-PSMNet trained weights for the Tri-Scene test, achieving an error rate of 15.1%, and directly used Tri-PSMNet trained weights for the InvTri-Scene test, achieving an error rate of 16.9%. The optimized universal framework, without the need for weight training, demonstrated higher accuracy and better adaptability to different scenes.