AirPose: Multi-View Fusion Network for Aerial 3D Human Pose and Shape Estimation

Overview

AirPose

AirPose: Multi-View Fusion Network for Aerial 3D Human Pose and Shape Estimation

iccv_teaser


Check the teaser video

This repository contains the code of AirPose a novel markerless 3D human motion capture (MoCap) system for unstructured, outdoor environments that uses a team of autonomous unmanned aerialvehicles (UAVs) with on-board RGB cameras and computation.

Please clone the repository with the following

git clone https://github.com/robot-perception-group/AirPose.git --recursive


Data can be freely accessed here. Please download the data, and untar it whenever necessary. Content details are following:

  • copenet_synthetic_data.tar.gz - Synthetic dataset
  • copenet_dji_real_data.tar.gz - real dataset
  • hmr_synthetic.tar.gz - Baseline pretrained checkpoint on synthetic data and pkl files for precalculated results
  • copenet_singleview_ckpt.zip - Baseline+Fullcam pretrained checkpoint on synthetic data and pkl files for precalculated results
  • muhmr_synthetic.tar.gz - Baseline+Multiview pretrained checkpoint on synthetic data and pkl files for precalculated results
  • copenet_twoview_synthetic_ckpt.tar.gz - AirPose pretrained checkpoint on synthetic data and pkl files for precalculated results
  • hmr_real_ckpt.zip - Baseline finetuned checkpoint on real data and pkl files for precalculated results
  • copenet_twoview_real_ckpt.zip - AirPose finetuned checkpoint on real data and pkl files for precalculated results
  • SMPLX_to_J14.pkl - Mapping from SMPLX joints to the 14 joints format of openpose. It is used in the method AirPose+.

The code was tested using Python 3.8.

SMPLX code in this repo is a modified version of the official SMLX implementation. Download the SMPLX model weights from here and run the following

# from the download location
unzip models_smplx_v1_1.zip -d models_smplx
unzip models_smplx/models/smplx/smplx_npz.zip -d used_models
rm models_smplx -r

Then copy the content of used_models (just created, with SMPLX_{MALE,FEMALE,NEUTRAL}.npz files) folder into your_path/AirPose/copenet/src/copenet/data/smplx/models/smplx.

You need to register before being able to download the weights.

Now, you may want to create a virtual environment. Please be sure your pip is updated.

Install the necessary requirements with pip install -r requirements.txt. If you don't have a cuda compatible device, change the device to cpu in copenet_real/src/copenet_real/config.py and copenet/src/copenet/config.py.

In those files (copenet_real/src/copenet_real/config.py and copenet/src/copenet/config.py) change LOCAL_DATA_DIR to /global/path/AirPose/copenet/src/copenet/data".

Check out this link to fix the runtime error RuntimeError: Subtraction, the - operator, with a bool tensor is not supported due to the Torchgeometry package.

Install the copenet and copenet_real packages in this repo

pip install -e copenet
pip install -e copenet_real

Download the head and hands indices files form here and place them in AirPose/copenet/src/copenet/data/smplx (MANO_SMPLX_vertex_ids.pkl and SMPL-X__FLAME_vertex_ids.npy).

Synthetic data training

The data to be used is copenet_synthetic_data.tar.gz (here)

Preprocess

To run the code of this repository you first need to preprocess the data using

# from AirPose folder
python copenet/src/copenet/scripts/prepare_aerialpeople_dataset.py /absolute/path/copenet_synthetic

And code can be run by the following (from AirPose/copenet folder):

python src/copenet/copenet_trainer.py --name=test_name --version=test_version --model=muhmr --datapath=/absolute/path/copenet_synthetic --log_dir=path/location/ --copenet_home=/absolute/path/AirPose/copenet --optional-params...

The datapath is the location of the training data.

--model specify the model type between [hmr, muhmr, copnet_singleview, copenet_twoview] which corresponds to the Baseline, Baseline+multi-view, Baseline+Fullcam and AirPose respectively.

Logs will be saved in $log_dir/$name/$version/

optional-params is to be substituted with the copenet_trainer available params as weights, lr..

Evaluation on the synthetic data

For model type [muhmr, copenet_twoview].

cd AirPose/copenet_real

python src/copenet_real/scripts/copenet_synth_res_compile.py "model type" "checkpoint Path" "/path to the dataset"

For model type [hmr, copenet_singleview], the provided checkpoint is trained with an older pytorch lightning version (<=1.2). If you want to use them, install pytorch-lightning<=1.2.

We provide the precalculated outputs on the syntehtic data using these checkpoints.

To generate the metrics, run

cd AirPose/copenet_real

python src/copenet_real/scripts/hmr_synth_res_compile.py "model type" "precalculated results directory Path" "/path to the dataset" "your_path/AirPose/copenet/src/copenet/data/smplx/models/smplx"

Fine-tuning on real dataset

The data to be used is copenet_dji_real_data.tar.gz(here).

Install the human body prior from here and download its pretrained weights (version 2) from here. Set the vposer_weights variable in the .../AirPose/copenet_real/src/copenet_real/config.py file to the absolute path of the downloaded weights (e.g. /home/user/Downloads/V02_05). If you do NOT have a GPU please change human_body_prior/tools/model_loader.py line 68 from state_dict = torch.load(trained_weigths_fname)['state_dict'] to state_dict = torch.load(trained_weigths_fname, map_location=torch.device('cpu'))['state_dict']

Note: for the hmr (Baseline) model pytorch-lightning<=1.2 is required. You might have to recheck requirements, or reinstall the requirements you can find in the main folder of this repo.

Code can be run by the following (from AirPose/copenet_real/ folder)

python src/copenet_real/copenet_trainer.py --name=test_name --version=test_version --model=hmr --datapath=path/location --log_dir=path/location/ --resume_from_checkpoint=/path/to/checkpoint --copenet_home=/absolute/path/AirPose/copenet --optional-params...

The datapath is the location of the training data.

--model specify the model type between [hmr, copenet_twoview] which corresponds to the Baseline, AirPose respectively.

The --resume_from_checkpoint is path to the pretrained checkpoint on the synthetic data.

Evaluation on real data

Install graphviz dependency with pip install graphviz in the same virtual environment.

Following code will generate the plots comparing the results of the baseline method, AirPose and AirPose+ on the real data.

This can be run from AirPose folder.

python copenet_real_data/scripts/bundle_adj.py "path_to_the_real_dataset" \\
"path_to_the_SMPLX_neutral_npz_file" \\
"path_to_vposer_folder" \\
"path_to_the_hmr_checkpoint_directory" \\
"path_to_the_airpose_precalculated_res_on_realdata_pkl" \\
"path_to_the_SMPLX_to_j14_mapping_pkl_file" \\
"type_of_data(train/test)" 

Note that:

  • The SMPLX_neutral_npz_file should be in your_path/AirPose/copenet/src/copenet/data/smplx/models/smplx.
  • The vposer_folder should be in the vposer_weights folder that you downloaded to finetune on the real data
  • The hmr checkpoint is either being generated by you or downloaded from here
  • The precalculated_res_on_realdata_pkl can be found within the same archive you downloaded above. More on how to compute them yourself below.
  • The SMPLX_to_j14_pkl can be found here.

The evaluation code above needs precalculated results on the real data which are provided with the dataset. If you want to calculate them yourself, run the following code and save the variable outputs in a pkl file when a breakpoint is hit. The pkl files provided with the data are generated in the same way. For AirPose

python copenet_real/src/copenet_real/scripts/copenet_real_res_compile.py "checkpoint Path" "/path to the dataset"

For Baseline

python copenet_real/src/copenet_real/scripts/hmr_real_res_compile.py "checkpoint Path" "/path to the dataset"

Testing the client-server synchronization mechanism

To this end you need to install ros-{melodic,noetic} in your pc (Ubuntu 18.04-20.04).

Please follow the instructions that you can find here

After that you need to install the following dependencies:

sudo add-apt-repository ppa:joseluisblancoc/mrpt-stable

Navigate to your catkin_ws folder (e.g. AirPose/catkin_ws) and run:

touch src/aircap/packages/optional/basler_image_capture/Grab/CATKIN_IGNORE
touch src/aircap/packages/optional/ptgrey_image_capture/Grab/CATKIN_IGNORE

this applies to ros-melodic

Firstly, checkout the AirPose branch ros-melodic.

Be sure to update the submodule (first command).

git submodule update 
sudo apt install libmrpt-dev mrpt-apps
cd /your/path/AirPose/catkin_ws
touch src/aircap/packages/3rdparty/mrpt_bridge/CATKIN_IGNORE
touch src/aircap/packages/3rdparty/pose_cov_ops/CATKIN_IGNORE
sudo apt install -y ros-melodic-octomap-msgs ros-melodic-cv-camera ros-melodic-marker-msgs ros-melodic-mrpt-msgs ros-melodic-octomap-ros ros-melodic-mrpt-bridge ros-melodic-mrpt1

this applies to ros-noetic

sudo apt install libmrpt-poses-dev libmrpt-obs-dev libmrpt-graphs-dev libmrpt-maps-dev libmrpt-slam-dev -y
sudo apt install -y ros-noetic-octomap-msgs ros-noetic-cv-camera ros-noetic-marker-msgs ros-noetic-mrpt-msgs ros-noetic-octomap-ros ros-noetic-mrpt2

Then you can run catkin_make from the catkin_ws folder to build the whole workspace.

To run the client-server architecture you need:

  • An image topic
  • A camera_info topic
  • A feedback topic with the region of interest information

To test the code you can do the following.

  • Download the dji rosbags that you can find here either from the test_bag or train_bag folders.
  • Download the checkpoint airpose_asv3_same_hparams_checkpoint.tar.gz from here

In separated terminals (with the workspace sourced) run:

  • roscore

  • rosparam set use_sim_time true

  • Launch the first client (i.e. the first "drone") with roslaunch airpose_client one_robot.launch host:=127.0.0.1 port:=9901 feedback_topic:=info img_topic:=camera/image_raw camera_info_topic:=camera/info robotID:=1 reproject:=false groundtruth:=true, with host you can change the server IP address, port must correspond, feedback_topic must contain the ROI and is of type neural_network_detector::NeuralNetworkFeedback, robotID should be either 1 or 2, reproject is used to avoid a reprojection to different intrisics parameters and groundtruth:=true is used to provide {min_x, max_x, min_y, max_y} in the ROI message (description below)

  • Launch the second client roslaunch airpose_client one_robot.launch host:=127.0.0.1 port:=9902 feedback_topic:=info img_topic:=camera/image_raw camera_info_topic:=camera/info robotID:=2 reproject:=false groundtruth:=true

  • Launch the servers, default IP 127.0.0.1

    Using the virtualenv/python3.8 installations with previous requirements installed

    • Firstly change folder to cd AirPose/catkin_ws/src/aircap/packages/flight/airpose_server
      • First server, run python server.py -p 9901 -m /path/to/the/file.ckpt, note that -p port needs to match the client_1 port
      • Second server, run python server.py -p 9902 -m /path/to/the/file.ckpt, note that -p port needs to match the client_2 port
  • To visualize the results you need to install some dependencies

    Using the virtualenv/python3.8 installations with previous requirements installed.

    • run pip install meshcat rospkg
    • Change folder to cd AirPose/catkin_ws/src/aircap/packages/flight/airpose_server and run pip install -e smplx.
    • The visualization node can then be run with python copenet_rosViz.py /machine_1/step3_pub /absolute/path/to/smplx/models or python copenet_rosViz.py /machine_2/step3_pub /absolute/path/to/smplx/models. The path is most likely /path/to/AirPose/copenet/src/copenet/data/smplx
  • You can either use the complete bag files with rosbag play d*_BGR.bag --clock --pause

    or create smaller (overlapping) bags using the split.zsh script that you find in both folders. This split will create 5 split from each bag. Afterwards, simply run rosbag play d*_split1.bag --clock --pause, where split{n-th} is the n-th split of the longer sequence. The splits have some overlap between them.

At this point you should be able to see play the rosbag in the way you prefer.

The published topics, for each machine, are:

  • machine_x/step1_pub, the results of the first step of the network, read by the other machine
  • machine_x/step2_pub, the results of the second step of the network, read by the other machine
  • machine_x/step3_pub, the final results of machine_x

The ROI message can be either used as "grountruth" box with the following structure:

ymin = ymin
ymax = ymax
ycenter = xmin
xcenter = xmax

Or as a more general box where you specify the center and the height of the box. In that case a 3:4 aspect ratio is considered.

ymin = ymin
ymax = ymax
xcenter = x_center_of_the_bb
ycenter = y_center_of_the_bb

You can also create your bag and provide your own data to the tool. To that end you can check the code available here that uses a csv with the needed information (image paths, bounding boxes, and camera info) to build the bags.

Note that this is no different than running the inference manually, except for the fact that this runs at 4FPS and has the synchronization procedure enabled as explained in the paper.

Owner
Robot Perception Group
Robot Perception Group is a reaearch group at the Perceiving Systems Department of the Max Planck Institue for Intelligent Systems, Tübingen, Germany
Robot Perception Group
MediaPipe is a an open-source framework from Google for building multimodal

MediaPipe is a an open-source framework from Google for building multimodal (eg. video, audio, any time series data), cross platform (i.e Android, iOS, web, edge devices) applied ML pipelines. It is

Bhavishya Pandit 3 Sep 30, 2022
Attempt at implementation of a simple GAN using Keras

Simple GAN This is my attempt to make a wrapper class for a GAN in keras which can be used to abstract the whole architecture process. Simple GAN Over

Deven96 7 May 23, 2019
Implementation of CaiT models in TensorFlow and ImageNet-1k checkpoints. Includes code for inference and fine-tuning.

CaiT-TF (Going deeper with Image Transformers) This repository provides TensorFlow / Keras implementations of different CaiT [1] variants from Touvron

Sayak Paul 9 Jun 26, 2022
LAMDA: Label Matching Deep Domain Adaptation

LAMDA: Label Matching Deep Domain Adaptation This is the implementation of the paper LAMDA: Label Matching Deep Domain Adaptation which has been accep

Tuan Nguyen 9 Sep 06, 2022
Code from PropMix, accepted at BMVC'21

PropMix: Hard Sample Filtering and Proportional MixUp for Learning with Noisy Labels This repository is the official implementation of Hard Sample Fil

6 Dec 21, 2022
PyTorch Kafka Dataset: A definition of a dataset to get training data from Kafka.

PyTorch Kafka Dataset: A definition of a dataset to get training data from Kafka.

ERTIS Research Group 7 Aug 01, 2022
RID-Noise: Towards Robust Inverse Design under Noisy Environments

This is code of RID-Noise. Reproduce RID-Noise Results Toy tasks Please refer to the notebook ridnoise.ipynb to view experiments on three toy tasks. B

Thyrix 2 Nov 23, 2022
Justmagic - Use a function as a method with this mystic script, like in Nim

justmagic Use a function as a method with this mystic script, like in Nim. Just

witer33 8 Oct 08, 2022
A data annotation pipeline to generate high-quality, large-scale speech datasets with machine pre-labeling and fully manual auditing.

About This repository provides data and code for the paper: Scalable Data Annotation Pipeline for High-Quality Large Speech Datasets Development (subm

Appen Repos 86 Dec 07, 2022
This is an official PyTorch implementation of Task-Adaptive Neural Network Search with Meta-Contrastive Learning (NeurIPS 2021, Spotlight).

NeurIPS 2021 (Spotlight): Task-Adaptive Neural Network Search with Meta-Contrastive Learning This is an official PyTorch implementation of Task-Adapti

Wonyong Jeong 15 Nov 21, 2022
B-cos Networks: Attention is All we Need for Interpretability

Convolutional Dynamic Alignment Networks for Interpretable Classifications M. Böhle, M. Fritz, B. Schiele. B-cos Networks: Alignment is All we Need fo

58 Dec 23, 2022
Official Pytorch implementation for 2021 ICCV paper "Learning Motion Priors for 4D Human Body Capture in 3D Scenes" and trained models / data

Learning Motion Priors for 4D Human Body Capture in 3D Scenes (LEMO) Official Pytorch implementation for 2021 ICCV (oral) paper "Learning Motion Prior

165 Dec 19, 2022
Official code release for 3DV 2021 paper Human Performance Capture from Monocular Video in the Wild.

Official code release for 3DV 2021 paper Human Performance Capture from Monocular Video in the Wild.

Chen Guo 58 Dec 24, 2022
SynNet - synthetic tree generation using neural networks

SynNet This repo contains the code and analysis scripts for our amortized approach to synthetic tree generation using neural networks. Our model can s

Wenhao Gao 60 Dec 29, 2022
This is official implementaion of paper "Token Shift Transformer for Video Classification".

This is official implementaion of paper "Token Shift Transformer for Video Classification". We achieve SOTA performance 80.40% on Kinetics-400 val. Paper link

VideoNet 60 Dec 30, 2022
Continuous Diffusion Graph Neural Network

We present Graph Neural Diffusion (GRAND) that approaches deep learning on graphs as a continuous diffusion process and treats Graph Neural Networks (GNNs) as discretisations of an underlying PDE.

Twitter Research 227 Jan 05, 2023
Stitch it in Time: GAN-Based Facial Editing of Real Videos

STIT - Stitch it in Time [Project Page] Stitch it in Time: GAN-Based Facial Edit

1.1k Jan 04, 2023
Companion code for the paper Theoretical characterization of uncertainty in high-dimensional linear classification

Companion code for the paper Theoretical characterization of uncertainty in high-dimensional linear classification Usage The required packages are lis

0 Feb 07, 2022
Official code release for ICCV 2021 paper SNARF: Differentiable Forward Skinning for Animating Non-rigid Neural Implicit Shapes.

Official code release for ICCV 2021 paper SNARF: Differentiable Forward Skinning for Animating Non-rigid Neural Implicit Shapes.

235 Dec 26, 2022
mmdetection version of TinyBenchmark.

introduction This project is an mmdetection version of TinyBenchmark. TODO list: add TinyPerson dataset and evaluation add crop and merge for image du

34 Aug 27, 2022