The invention discloses a method based on RGB-. According to the indoor scene three-dimensional reconstruction method of the D image, a semantic segmentation result is used for repairing a depth imagecavity, object contour and category information is provided for three-dimensional reconstruction, and the shape and appearance of an object are obtained according to priori knowledge, so that more accurate data is provided for three-dimensional reconstruction. Three-dimensional reconstruction provides three-dimensional space information for semantic segmentation, and the problem of false segmentation caused by object overlapping, illumination influence and the like in two-dimensional image segmentation is solved. Multi-level camera pose estimation and sparse feature matching are used for providing a rough estimation pose, then a precise camera pose is obtained through a dense geometric and photometric optimization method, and a more precise camera pose is provided for model reconstruction. In the reconstruction process, each frame is locally optimized, meanwhile, a key frame mechanism is added, global optimization and closed-loop detection are established, constraint is established for spatial points corresponding to pixels of the key frames, error accumulation is effectively inhibited, the pose of the camera is further optimized, and the precision of a reconstruction result is improved.