LUT-DM-filters

Starting with the implementation outlined in Diffpose, we streamlined the model architecture by removing the context encoder and GMM components and we customized it to handle only 12 human keypoints.

MICRO-BENCHMARKING

Look at lut_analysis.ipynb

LUT-based real-time filtering

Download the dataset of Total Capture here.

Apply noise to the ground truth dataset, creating a hierarchical structure as depicted below:

data
└── total_capture
    ├── noisy
    └── gt

LUT-DM / SMART-DM

To run the LUT-DM or SMART-DM or ORACLE scripts, use the following command-line format:

python <file_name> --path <path> --gt_path <gt_path> 

Where:

• <file_name>: Name of the Python script to execute.
  • Options: lut_launch_name.py, smart_launch_name.py, oracle_bench_name.py
• <path>: path to the folder containing noisy data.
• <gt_path>: path to the folder containing ground truth data.

To run the DDPM or DDIM scripts, use the following command-line format:

python ddpm_name.py --path <path> --gt_path <gt_path> --mode <mode> 

Where:

• <path>: path to the folder containing noisy data.
• <gt_path>: path to the folder containing ground truth data.
• <mode>: Type of noise to apply to the dataset.
  • Options: ddpm, ddim

Examples:

python ddpm_name.py --path data/total_capture/noisy/ --gt_path data/total_capture/gt/ --mode ddpm

Note: In our implementation we use common and models folder from Diffpose