There are two ways of pose calculation: SLAM and SFM. SLAM calculates the relative position of each frame of image. If you only use pictures to calculate, you can’t get absolute scale information, and there will be a scaling ratio between the reconstructed and real objects. If you want to get the absolute size, you may need to add some auxiliary information. For example, use an IMU or a binocular camera, or place an object of known size in the scene.
The difference between SLAM and SFM:
SLAM is suitable for real-time performance. It requires the input to be a video sequence, and there is temporal continuity between frames.
SFM is suitable for offline tasks. Its input can be found on the Internet, even pictures taken by different devices, or video sequences.
The MVS performs dense reconstruction after obtaining the pose information.
Traditional methods of 3D reconstruction based on MVS:
1. Input image and pose. Some other information can be added, such as sparse points or depth ranges.
2. Dense reconstruction. This step is equivalent to obtaining the 3D information in the scene. If this step is accurate, the subsequent reconstruction will be more accurate.
3. Point cloud fusion of depth map.
4. Surface reconstruction/mesh reconstruction.
5. Grid optimization. Improve details.
6. Texture map. The higher the image resolution, the higher the quality of the texture map.