Abstract: 3D reconstruction is a key problem in computer vision.With the advances of 3D reconstruction techniques,people more and more realized that under the tradition paradigm purely from images to 3D structure,it is nearly impossible to recover the scene structure in a semantic sense no matter how deliberate design and optimization are employed.To this end,a semantic interaction based scene reconstruction approach is explored where automatically reconstructed noisy points or lines are upgraded through natural language understanding.The feasibility of our approach is illustrated by two simple experiments.