基于边缘辅助极线Transformer的多视角场景重建

基于深度学习的多视角立体几何(MVS)旨在通过多个视图重建出稠密的3维场景。然而现有的方法通常设计复杂的2D网络模块来学习代价体聚合的跨视角可见性，忽略了跨视角2维上下文特征在3D深度方向的一致性假设。此外，基于多阶段的深度推断方法仍需要较高的深度采样率，并且在静态或预先设定的范围内采样深度值，容易在物体边界以及光照遮挡等区域产生错误的深度推断。为了缓解这些问题，该文提出一种基于边缘辅助极线Transformer的密集深度推断模型。与现有工作相比，具体改进如下：将深度回归转换为多深度值分类进行求解，在有限的深度采样率和GPU占用下保证了推断精度；设计一种极线Transformer模块提高跨视角代价体聚合的可靠性，并引入边缘检测分支约束边缘特征在极线方向的一致性；为了提高弱纹理区域的精度，设计了基于概率成本体积的动态深度范围采样机制。与主流的方法在公开的数据集上进行了综合对比，实验结果表明所提模型能够在有限的显存占用下重建出稠密准确的3D场景。特别地，相比于Cas-MVSNet，所提模型的显存占用降低了35%，深度采样率降低约50%，DTU数据集的综合误差从0.355降低至0.325。

Multiview Scene Reconstruction Based on Edge Assisted Epipolar Transformer

Learning-based Multiple-View Stereo (MVS) aims to reconstruct dense 3D scene representation. However, previous methods utilize additional 2D network modules to learn the cross view visibility for cost aggregation, ignoring the consistency assumption of 2D contextual features in the 3D depth direction. In addition, the current multi-stage depth inference model still requires a high depth sampling rate, and depth hypothesis is sampled within static and preset depth range, which is prone to generate errorneous depth inference in the object boundary and occluded area. To alleviate these problems, a multi-view stereo network based on edge assisted epipolar Transformer is proposed. The improvements of this work over the state of the art are as: Depth regression is replaced by the multi-depth hypotheses classification to ensure the accuracy with limited depth sampling rate and GPU consumption. Epipolar Transformer block is developed for reliable cross view cost aggregation, and edge detection branch is designed to constrain the consistency of edge features in the epipolar direction. A dynamic depth range sampling mechanism based on probabilistic cost volume is applied to improve the accuracy of uncertain areas. Comprehensive comparisons with the state of the art are conducted on public benchmarks, which indicate that the proposed method can reconstruct dense scene representations with limited memory bottleblock. Specifically, compared with Cas-MVSNet, the memory consumption is reducted by 35%, the depth sampling rate is reduced by about 50%, and the overall error on DTU datasets is reduced from 0.355 to 0.325.