Privacy Scan

Due to privacy concerns, scanners that use infrared light (such as Kinect and RealSense) are sensitive to dust and hair, and the surfaces of the models they create are often very rough, making them unsuitable for direct viewing by the user. Furthermore, survey results show that users tend not to want their own faces to be transferred to their scanned versions. In response to requests for models to be able to view their own faces, we developed a process for reconstructing scanned models.

point:

These processes are necessary to make the model look like a human being rather than just clay.

• aesthetic

• Face, limb reconstruction

• material

• Rigging

After predefining the template and performing generality fitting iterative retopology with TG3D body landmark constraints, the mesh distribution is good except for the wrists and head.

Template mapping results:

Gray is raw data, and skin color is the result of template mapping.

Data Flaw:

Issue 1. Marubozu forecast return:

The presence of noise and hair can easily lead to mismatched geometric fitting, especially for people with long bangs, where the face mesh is projected onto the hair, causing distortions and negatively impacting the machine learning model later.

Collect data from studies around the world and perform rigid implantation of template heads with a range of confidence intervals for head and hand sizes.

Issue 2. Filtering fraudulent data and reducing the amount outside the scan range

Statistical reconstruction of head size from scan data.

result:

From the successfully reconstructed images, various information such as contours, normals, depth, and measurements are extracted, and machine learning is used to support scannerless scanning. Since the images are extracted from the actual model, they can also be sent to Stable Diffusion's ControlNet for AI image generation.