基于细粒度文本特征融合的跨模态单目深度估计模型OA

In the field of monocular depth estimation for complex scenes,a key problem is the occurrence of depth discontinuities and blurred edges caused by irregular objects such as people or animals.To address this issue,a model named Multiple Fine-Grained Textual Features Fusion for Monocular Depth Estimation(MFGT-Depth)was designed.The Gaussian Mixture Model(GMM)was used to probabilistically model latent semantic hier-archy and complex relationships among descriptions.For efficient cross-modal alignment,the cross-modal module then adaptively fused the text features and the image features ex-tracted by the image encoder.Experiments show that MFGT-Depth achieves significant improvements in major metrics on training datasets,reducing RMSE by 4.8%on the KITTI dataset,and exhibits strong generalization with significantly improved accuracy and robustness for depth estimation in scenes with irregular objects relative to Depth Any-thing V2.