Skip to content

Latest commit

 

History

History
50 lines (37 loc) · 3.7 KB

README.md

File metadata and controls

50 lines (37 loc) · 3.7 KB
Raw

Coordinate-Wise Monotonic Transformations Enable Privacy-Preserving Age Estimation with 3D Face Point Cloud

by Xinyu Yang, Runhan Li, Xindi Yang, Yong Zhou, Yi Liu & Jing-Dong J. Han*

Created in NightCafe with prompts "a cyberpunk neural network as background with a 3D meshed construction toy-like face without body in the center and robotic arms in the periphery touching the face with flash and flares. Picture in no more than 6 colors with the "viridis" palette."

Summary of the findings

  1. Registration-free and accurate age estimation with 3D face point cloud input.
  2. Coordinate-wise Monotonic Transformations (CMTs) that distort face shape without compromising age-related information.
  3. CMTs invalidate a 3D face verification model trained with face shape, disentangling age and identity information of the face.

Age estimation with 3D face point cloud

Use scripts in FPCT-age. Put your data split record (simply in txt format) in DataSplit. Modify config.yaml in config for your training specifications. Run script_main.py with hydra-style parameter passing (https://hydra.cc/docs/intro/) to train. Modify batch_test_config.yaml in config for your inference specifications. Run batch_test_main.py with hydra-style parameter passing to evaluate.

e.g.

python script_main.py ozername=SGD basic_learning_rate=5e-5 # train
python batch_test_main.py dataset=path/to/your/data foldername=path/to/save/results # infer and evaluate

Face verification with 3D face point cloud

Use scripts in FPCT-ID. Put your data split record (simply in txt format) in DataSplit. Modify config.yaml in config for your training specifications. Run script_main.py with hydra-style parameter passing to train. Modify batch_inter_test_config.yaml in config for your inference specifications. Run batch_inter_test_main.py with hydra-style parameter passing to infer. Run GenROC.py to get ROC and AUC using a pair of training data inference and test data inference.

e.g.

python script_main.py ozername=SGD basic_learning_rate=5e-5 # train
python batch_inter_test_main.py dataset=path/to/your/data foldername=path/to/save/results # infer
python GenROC.py # evaluate

VisualCMT

Download and unzip VisualCMT_compressed.zip, then double click VisualCMT.exe to use. It is an interactive software for Coordinate-wise Monotonic Transformations (CMTs). The software takes in comma-separated value (CSV) formatted data with X, Y, and Z coordinates in the first three columns, and R, G, B color components in the last three columns. The software automatically normalizes the X, Y, and Z coordinates to the range of [-1, 1], while RGB values are expected to be provided in decimal magnitudes. Mouse drag and wheel spin are supported in the input and output view windows. Six CMTs are available including Coordinate-wise Rank Transformation (CRT).

Input data structure example:

    x         y         z        xn        yn        zn         r         g         b
0 -0.784874 -0.244051 -0.536802 -0.455303 -0.903667 -0.458548  0.053385  0.033854  0.020833
1 -0.779600 -0.246530 -0.537601 -0.384419 -0.919449 -0.497628  0.057292  0.037760  0.023438
2 -0.795638 -0.237591 -0.536667 -0.584383 -0.838801 -0.435541  0.061198  0.031250  0.022135
3 -0.789381 -0.241684 -0.536477 -0.541382 -0.862164 -0.443775  0.052083  0.027344  0.020833
4 -0.787205 -0.242815 -0.536962 -0.496568 -0.885021 -0.457656  0.057292  0.033854  0.022135

VisualCMT screenshot:

The software was packed using PyInstaller and mannually reduced volume in a Win11 system, please report issue if the software fail to run in your system.