Data preparation for Real-to-Sim in Customized Scene

[HiFunRobo]

Prepare the data:
Place the images in the folder /path/to/input/images.
Place the image for alignment in /path/to/input/align_image.png.

what are the differences between images and align_image.png.? how many images should I prepare, and at what angles should these images be taken?
Please give some more detailed instructions, and it would be best if you can include a demo data folder for reference.

Thank you!

[xshenhan]

Thank you for pointing this out. The two inputs serve different purposes:

/path/to/input/images contains the multi-view images used to reconstruct the background, including the Gaussian representation and the background mesh. These images should mainly capture the static background scene without the robot, so that the reconstruction does not contain robot artifacts.

/path/to/input/align_image.png is used only for aligning the robot with the reconstructed background. This image should contain both the robot and the background in the same view. In most cases, one such image is enough.

For /path/to/input/images, we recommend preparing multiple images taken from different viewpoints around the scene. The exact number depends on the scene size and complexity, but generally 30–100 images are sufficient for a small tabletop or room-scale scene. The images should have enough overlap with each other, cover the scene from different angles and heights, and avoid severe motion blur or large lighting changes. A good rule of thumb is to move the camera smoothly around the scene while keeping the target area visible, similar to standard COLMAP / 3D Gaussian Splatting data collection.

For align_image.png, please choose one clear image where the robot and the background are both visible. This image is not used for reconstruction; it is only used to estimate the relative alignment between the robot and the reconstructed scene.