Skip to content

Latest commit

 

History

History
105 lines (75 loc) · 2.87 KB

README.md

File metadata and controls

105 lines (75 loc) · 2.87 KB
Raw

Structure

The detailed structure is as follows:

- afford-motion/
  - body_models/
    - smplx/
      - SMPLX_NEUTRAL.npz
      - ...
  - data/
    - custom/
    - eval/
    - H3D/
    - HUMANISE/
    - HumanML3D/
    - PROX/
    - POINTTRANS_C_N8192_E300/
    - Mean_Std_*
    - ...
  - outputs/
    - CDM-Perceiver-H3D/                # pre-trained ADM model on original HumanML3D dataset
    - CDM-Perceiver-HUMANISE-step200k/  # pre-trained ADM model on HUMANISE dataset
    - CDM-Perceiver-ALL/                # pre-trained ADM model on all datasets for novel set evaluation
    - CMDM-Enc-H3D-mixtrain0.5/         # pre-trained CMDM model on original HumanML3D dataset
    - CMDM-Enc-HUMANISE-step400k        # pre-trained CMDM model on HUMANISE dataset
    - CMDM-Enc-ALL/                     # pre-trained CMDM model on all datasets for novel set evaluation
  - configs/
  - datasets/
  - ...

Process Data

The following process is to prepare the data for training model using HumanML3D, HUMANISE, and PROX together, corresponding to the evaluation on Novel Evaluation Set.

We also use the processed HUMANISE data to train and evaluate on HUMANISE benchmark.

1. Process motion of each dataset

HumanML3D

After download the AMASS dataset (both SMPL+H and SMPL-X versions), run the following commands:

python prepare/process.py --dataset HumanML3D --data_dir ${YOUR_PATH}/amass/smplx_neutral

The SMPL+H data should be put into ${YOUR_PATH}/amass/smplh. The SMPL-X data should be put into ${YOUR_PATH}/amass/smplx_neutral.

Copy the original texts.zip into ./data/HumanML3D folder and unzip it.

Process HUMANISE

After download the HUMANISE dataset, run the following commands:

python prepare/process.py --dataset HUMANISE --data_dir ${YOUR_PATH}/HUMANISE

The align_data_release folder and pure_motion folder should both be put into ${YOUR_PATH}/HUMANISE.

Process PROX

We use the refined version of PROX's per-frame SMPL-X parameters from LEMO. Please download the the PROX scene and cam2world data and LEMO motion data.

Move the PROX scene and cam2world data into ./data/PROX folder.

Then run the following commands to process LEMO (motion):

python prepare/process.py --dataset PROX --data_dir ${YOUR_PATH}/LEMO/PROX_temporal/PROX_temporal/PROXD_temp

2. Convert SMPL-X to vectorized representations (joint positions)

python prepare/smplx_to_vec.py --dataset ${DATASET}
# e.g., python prepare/smplx_to_vec.py --dataset PROX

3. Process the scene point cloud

python prepare/process_scene.py

4. Generate Contact Data

python prepare/generate_contact_data.py --random_segment

5. Re-split the dataset

python prepare/split.py

Others

  • Generate target mask for HUMANISE dataset (used in evaluation):
python prepare/generate_target_object_mask.py