WT-MVSNet: Window-basedTransformersfor Multi-viewStereo

Arecenteffort toperform attention-based matching alongtheepipolar linesofsourceimages [32],suffersinstead from sensitivity to inaccurate camera pose and calibration, which can in turn results to erroneous matching. Another key step in contemporary learned MVS methods is the regularization of cost volume, generated by stacking cost maps associated with respective depth hypotheses.