This is the official implementation of our work presented at CVPR 2024, titled Multiple View Geometry Transformers for 3D Human Pose Estimation.
We explicitly introduce multi-view geometric modules into an end-to-end Transformers architecture for 3D human pose estimation, resulting in remarkable generalization performance across various camera settings.
We employ a coarse-to-fine query refinement process, initially randomly sampling coarse 3D queries in the areas. Subsequently, we project these coarse 3D poses onto 2D images and aggregate features using Transformer attentions to update the projected points. Finally, we use triangulation to recover more accurate 3D poses. This process is iteratively repeated in an end-to-end differentiable manner by Transformer decoder layers.
@inproceedings{liao2024multiple,
title={Multiple View Geometry Transformers for 3D Human Pose Estimation},
author={Liao, Ziwei and Zhu, Jialiang and Wang, Chunyu and Hu, Han and Waslander, Steven L},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
pages={708--717},
year={2024}
}
Clone the repository:
git clone git@github.com:XunshanMan/MVGFormer.git
Create a conda environment:
conda create -n mvgformer python==3.10
conda activate mvgformer
Install mmcv carefully with cuda version:
pip install -U openmim
mim install mmcv-full
Install the dependencies:
pip install -r requirements.txt
Compile the attention modules. Please carefully set your CUDA_HOME=/usr/local/cuda-*/, where * is your cuda version:
cd ./lib/models/ops
CUDA_HOME=/usr/local/cuda-11.0/ python setup.py build install
Consider the project root directory as ${POSE_ROOT}.
CMU Panoptic dataset. Please download the CMU Panoptic Dataset following VoxelPose as below:
-
Download the dataset by following the instructions in panoptic-toolbox and extract them under
${POSE_ROOT}/data/panoptic/ -
You can only download those sequences you need. You can also just download a subset of camera views by specifying the number of views (HD_Video_Number) and changing the camera order in
./scripts/getData.sh. The sequences and camera views used in our project can be obtained from here, and theTable A.1in the Supplementary Materials of our paper. -
Note that we only use HD videos, calibration data, and 3D Body Keypoint in the codes. You can comment out other irrelevant codes such as downloading 3D Face data in
./scripts/getData.sh.
Feature Backbone. Please download the PoseResNet-50 pre-trained model and place it at ${POSE_ROOT}/models/pose_resnet50_panoptic.pth.tar (ResNet-50 pretrained on COCO dataset and finetuned jointly on Panoptic dataset and MPII).
The structure should look like this:
${POSE_ROOT}
|-- models
| |-- pose_resnet50_panoptic.pth.tar
|-- data
|-- panoptic
|-- 16060224_haggling1
| |-- hdImgs
| |-- hdvideos
| |-- hdPose3d_stage1_coco19
| |-- calibration_160224_haggling1.json
|-- 160226_haggling1
|-- ...
Shelf/Campus Dataset. Please follow VoxelPose to download the Shelf/Campus Dataset.
We train on five camera views from the camera arrangement of CMU0. We trained our models on 8 GPUs and batch_size=1 for each GPU. The evaluation result will be printed after every epoch, the best result can be found in the log.
python -m torch.distributed.launch --nproc_per_node=8 --use_env run/train_3d.py --cfg configs/panoptic/knn5-lr4-q1024-g8.yaml
The trained model checkpoints can be downloaded as below. Please put the weights under models/.
| Datasets | AP25 | MPJPE | pth |
|---|---|---|---|
| Panoptic | 92.3 | 16.0 | here |
Following Section 2.1, please change the parameters in the config file, or pass through extra parameters, such as DECODER.num_decoder_layers=4, DECODER.num_instance=1024, TRAIN.END_EPOCH=100 to run the ablation experiments.
We train our model on the CMU panoptic dataset, with the camera arrangement of CMU0. We can infer the model on several out-of-domain settings, including Change Camera Numbers, Change Camera Arrangements, and Change Dataset without finetuning.
For CMU-panoptic dataset, you can find ready-made configuration files below configs/panoptic/generalization. EXP_TYPE can be CMU0ex3,CMU0ex4,CMU0ex6,CMU0ex7,CMU1,CMU2,CMU3,CMU4.
python run/validate_3d.py --cfg configs/panoptic/generalization/{EXP_TYPE}.yaml --model_path models/mvgformer_q1024_model.pth.tar
You can also configure experiments with extra parameters:
Change Camera Numbers. The camera number can be3,4,5,6,7:
python run/validate_3d.py --cfg configs/panoptic/knn5-lr4-q1024.yaml --model_path models/mvgformer_q1024_model.pth.tar DATASET.TEST_CAM_SEQ='CMU0ex' DATASET.CAMERA_NUM=7
Change Camera Arrangements. The camera arrangements can beCMU0,CMU1,CMU2,CMU3,CMU4. Please see here for the detail of each arrangements:
python run/validate_3d.py --cfg configs/panoptic/knn5-lr4-q1024.yaml --model_path models/mvgformer_q1024_model.pth.tar DATASET.TEST_CAM_SEQ=CMU1
Change Dataset without finetuning. The dataset can beShelf,Campus.
python run/validate_3d.py --cfg XXXX --model_path models/mvgformer_q1024_model.pth.tar --dataset Shelf/Campus
This experiment aims to test the in-domain performance of our model. It is trained on the camera arrangement of CMU0 on the CMU panoptic dataset, and also infer with the same camera arrangement.
Infer with single GPU with BATCH_SIZE=1:
python run/validate_3d.py --cfg configs/panoptic/knn5-lr4-q1024.yaml --model_path models/mvgformer_q1024_model.pth.tar TEST.BATCH_SIZE=1
This repository is licensed under Apache-2.0. For commercial use, please reach out to the authors.
Our work is based on the codebases of MvP and VoxelPose. We deeply appreciate the authors for their invaluable contributions.