This repos contains the implementation of a deep learning pipeline that generates a 3D human avatar with 22 body measurements from just a pair of input front side human images within seconds.
Generally, two input images will be passed to a deep learning model to estimate the human silhouettes, which are then stacked together and passed to another deep learning model that predicts body parameters of a statistical human model. These body parameters will be then optimized to fit better to body keypoints landmarks and the silhouettes. Finally, a human mesh is calculated and the 22 body measurements are extracted.
Following this instruction to quickly bring up the whole pipeline.
The system also comes with a web portal for testing the main features of the pipeline.
- Manage and store multiple subjects in a SQL database.
- Predict 3D human shapes for subject
- Calculate measurements on the predicted 3D body shapes.
- Support compare the measurement error given the groundtruth measurements.
To bring up the web portal with the pre-trained models (stored in google drive), please follows the instructions