DepthMamba:多尺度VisionMamba架构的单目深度估计  

DepthMamba:multi-scale VisionMamba architecture for monocular depth estimation

作　　者：徐志斌 张孙杰 Xu Zhibin;Zhang Sunjie(School of Optical-Electrical&Computer Engineering,University of Shanghai for Science&Technology,Shanghai 200093,China)

机构地区：[1]上海理工大学光电信息与计算机工程学院,上海200093

出　　处：《计算机应用研究》2025年第3期944-948,共5页Application Research of Computers

基　　金：国家自然科学基金资助项目(61603255);上海市晨光计划资助项目(18CG52)。

摘　　要：在单目深度估计领域,虽然基于CNN和Transformer的模型已经得到了广泛的研究,但是CNN全局特征提取不足,Transformer则具有二次计算复杂性。为了克服这些限制,提出了一种用于单目深度估计的端到端模型,命名为DepthMamba。该模型能够高效地捕捉全局信息并减少计算负担。具体地,该方法引入了视觉状态空间(VSS)模块构建编码器-解码器架构,以提高模型提取多尺度信息和全局信息的能力。此外,还设计了MLPBins深度预测模块,旨在优化深度图的平滑性和整洁性。最后在室内场景NYU_Depth V2数据集和室外场景KITTI数据集上进行了综合实验,实验结果表明:与基于视觉Transformer架构的Depthformer相比,该方法网络参数量减少了27.75%,RMSE分别减少了6.09%和2.63%,验证了算法的高效性和优越性。In the field of monocular depth estimation,researchers have extensively studied models based on CNN and Transformer.However,CNN struggle with inadequate extraction of global features,while Transformer exhibit quadratic computational complexity.To overcome these limitations,this paper proposed an end-to-end model DepthMamba for monocular depth estimation.The model was able to capture global information efficiently and reduce the computational burden.Specifically,the method introduced a visual state space(VSS)module to construct an encoder-decoder architecture to improve the model s ability to extract multi-scale information and global information.Additionally,this paper designed an MLPBins depth prediction module to ensure smoother and cleaner generated depth maps.This paper conducted comprehensive experiments on indoor scenes using the NYU_Depth V2 dataset and outdoor scenes using the KITTI dataset.Compared with the Depthformer architecture based on vision Transformer,this method reduced network parameters by 27.75%and decreases the RMSE by 6.09%and 2.63%,respectively,which validates the algorithm s efficiency and superiority.

关 键 词：单目深度估计 Vmamba Bins深度预测 状态空间模型 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]