Skip to content

Latest commit

 

History

History
261 lines (203 loc) · 9.98 KB

README.md

File metadata and controls

261 lines (203 loc) · 9.98 KB
Raw

Toward Reliable Human Pose Forecasting with Uncertainty

UnPOSed is an open-source toolbox for prediction/forecasting a sequence of human pose given an observed pose sequence, implemented in PyTorch.

IEEE Robotics and Automation Letters, 2024

[paper] [livedemo]

Input pose
observation

Output pose and ground-truth
prediction

Overview

The main parts of the library are as follows:

unposed
├── api
|   ├── preprocess.py                   -- script to run the preprocessor module
|   ├── train.py                        -- script to train the models, runs factory.trainer.py
│   ├── evaluate.py                     -- script to evaluate the models, runs factory.evaluator.py
|   └── generate_final_output.py        -- script to generate and save the outputs of the models, runs factory.output_generator.py
├── models                    
|   ├── st_trans/ST_Trans.py
|   ├── pgbig/pgbig.py
│   ├── history_repeats_itself/history_repeats_itself.py
│   ├── sts_gcn/sts_gcn.py
|   ├── zero_vel.py
│   ├── msr_gcn/msrgcn.py
|   ├── potr/potr.py
|   ├── pv_lstm.py
│   ├── disentangled.py
|   ├── derpof.py
|   ├── deposit/deposit.py
|   ├── ...
├── losses
|   ├── pua_loss.py
│   ├── mpjpe.py           
|   ├── ...

The library has four important APIs to

  • preprocess data
  • train the model
  • evaluate a model quantitatively
  • generate their outputs

The details of how to use these API are described below. Two other important directories are models and losses. In these two directories, you can add any desired model and loss function and leverage all predefined functions of the library to train and test and compare in a fair manner.

Please check other directories (optimizers, mmetrics, schedulers, visualization, utils, etc.) for more abilites.

Getting Started

To get started as quickly as possible, follow the instructions in this section. This should allow you train a model from scratch, evaluate your pretrained models, and produce basic visualizations.

Dependencies

Make sure you have the following dependencies installed before proceeding:

  • Python 3.7+ distribution
  • PyTorch >= 1.7.0
  • CUDA >= 10.0.0
  • pip >= 21.3.1

Virtualenv

You can create and activate virtual environment like below:

pip install --upgrade virtualenv

virtualenv -p python3.7 <venvname>  

source <venvname>/bin/activate  

pip install --upgrade pip

Requirements

You can install all the required packages with:

pip install -r requirements.txt

Before moving forward, you need to install Hydra and know its basic functions to run different modules and APIs.

Hydra

Hydra is a framework for elegantly configuring complex applications with hierarchical structure. For more information about Hydra, read their official page documentation.

In order to have a better structure and understanding of our arguments, we use Hydra to dynamically create a hierarchical configuration by composition and override it through config files and the command line. If you have any issues and errors install hydra like below:

pip install hydra-core --upgrade

for more information about Hydra and modules please visit here

MLFlow

We use MLflow in this library for tracking the training process. The features provided by MLFlow help users track their training process, set up experiments with multiple runs and compare runs with each other. Its clean, and organized UI helps users to better understand and track what the experiments. See more from MLFlow here.

This part is not obligatory but you can use MLFlow by running the command below in the folder containing mlruns folder to track the training processes.

mlflow ui

Datasets

We currently Support the following datasets:

  • Human3.6M in exponential map can be downloaded from here.
  • AMASS from their official website..
  • 3DPW from their official website.

Please download the datasets and put them in a their specific folder. We will refer to this folder as $DATASET_PATH in the following sections.

Models

We've tested the following models:

Adding a Model

To add a new model, you need to follow the below steps:

  • add the model file or files in the model directory
  • add the model reference to the models.__init__.py
  • add the model's required parameters to the configs/hydra/models. This step is necessary even if you don't have additional parameters
  • if your model has new loss function which is not implemented in the library, you can add your loss function to the losses folder.

Adding a Metric

To add a new metric, you need to follow the below steps:

  • implement your metric function in the metrics.pose_metrics.py file
  • add the model reference to the metrics.__init__.py
  • add your metric to the configs/hydra/metrics.yml

Preprocessing

We need to create clean static files to enhance dataloader and speed-up other parts. To fulfill mentioned purpose, put the data in DATASET_PATH and run preprocessing api called preprocess like below:

Example:

python -m api.preprocess \
    dataset=human3.6m \
    annotated_data_path=$DATASET_PATH \
    data_type=test

See here for more details about preprocessing arguments. This process should be repeated for training, validation and test set. This is a one-time use api and later you just use the saved jsonl files.

Training

Given the preprocessed data, train models from scratch:

python -m api.train model=st_trans \
    train_dataset=$DATASET_TRAIN_PATH \
    valid_dataset=$DATASET_VALIDATION_PATH \
    obs_frames_num=50 \
    pred_frames_num=25 \
    model.loss.nT=25 \
    model.pre_post_process=human3.6m \
    model.n_major_joints=22 \
    model.loss.nJ=32

DATASET_TRAIN_PATH and DATASET_VALIDATION_PATH refer to the preprocessed json files.

NOTE: You can see more commands for training models here.

Provide validation_dataset to adjust learning-rate and report metrics on validation-dataset as well.

See here for more details about training arguments.

Evaluation

You can evaluate any pretrained model with:

python -m api.evaluate model=st_trans \
          dataset=$DATASET_TEST_PATH \
          load_path=$MODEL_CHECKPOINT \
          obs_frames_num=10 \
          pred_frames_num=25 \
          data.is_h36_testing=true

in case of non-trainable model, run:

python -m api.evaluate model=zero_vel \
          dataset=$DATASET_TEST_PATH \
          obs_frames_num=10 \
          pred_frames_num=25 \
          data.is_h36_testing=true

See here for more details about evaluation arguments.

Epistemic Uncertainty

You can also evaluate the epistemic uncertainty of the models using the approach presented in the paper. For the ease of use, we have provided the trained EpU model in the release section. Using the trained EpU model, one can evaluate the epistemic uncertainty of the pose prediction model:

python -m api.evaluate model=st_trans \
          dataset=$DATASET_TEST_PATH \
          load_path=$MODEL_CHECKPOINT \
          obs_frames_num=10 \
          pred_frames_num=25 \
          data.is_testing=true \
          data.is_h36_testing=true \
          dataset_name=human3.6m   \
          eval_epu=true \
          epu_model_path=$EPU_MODEL

Generating Outputs

Generate and save the predicted future poses:

python -m api.generate_final_output model=st_trans \
          dataset=$DATASET_PATH \
          load_path=$MODEL_CHECKPOINT \
          obs_frames_num=10 \
          pred_frames_num=25 \
          data.is_h36_testing=true \
          save_dir=$OUTPUT_PATH

See here for more details about prediction arguments.

Visualization

Launch an interactive web app to visualize model outputs by running:

python -m visualization.server

For more details, visit visualization/README.md.

Work in Progress

This repository is being updated so please stay tuned!

@article{saadatnejad2024unposed,
  title={Toward reliable human pose forecasting with uncertainty},
  author={Saadatnejad, Saeed and Mirmohammadi, Mehrshad and Daghyani, Matin and Saremi, Parham and Benisi, Yashar Zoroofchi and Alimohammadi, Amirhossein and Tehraninasab, Zahra and Mordan, Taylor and Alahi, Alexandre},
  journal={IEEE Robotics and Automation Letters},
  year={2024},
  publisher={IEEE}
}