一种弱纹理目标立体匹配网络

A Stereo Matching Network for Weak Texture Objects

摘要:

鉴于传统深度估计方法在高分辨率图像下存在特征提取不够充分、图像信息获取不完整、受限于局部信息或特定类型的特征提取等问题,为此提出一种面向全局特征的Transformer立体匹配网络。该网络采用编码器-解码器的端到端架构,使用多头注意力机制,允许模型在不同子空间中关注不同的特征,从而提高建模能力。模型将自注意力机制和特征重构窗口相结合,能够提高特征的表征能力,弥补局部特征不足问题,减少计算负担的同时有效应对Transformer架构通常伴随的高计算复杂度问题,确保模型的注意力计算保持在线性复杂度范围内。在Scene Flow、KITTI-2015数据集上分别进行实验,指标获得显著提升,通过对比实验验证模型的有效性和正确性。

Abstract:

Given that traditional depth estimation methods have problems such as insufficient feature extraction, incomplete image information acquisition, and limitations on local information or specific types of feature extraction in high-resolution images, a transformer stereo matching network for global features is proposed. The network adopts an end-to-end architecture of encoder decoder, using a multi head attention mechanism to allow models to focus on different features in different sub-spaces, thereby improving modeling capabilities. The model combines self attention mechanism and feature reconstruction window to improve feature representation ability, compensate for local feature deficiencies, reduce computational burden, and effectively address the high computational complexity typically associated with Transformer architectures, Ensure that the attention calculation of the model remains within the linear complexity range; Experiments were conducted on the Scene Flow and KITTI-2015 datasets, and the indicators were significantly improved; The effectiveness and correctness of the model were verified through comparative experiments.