Code repository for the paper:
Score-Guided Diffusion for 3D Human Recovery
Anastasis Stathopoulos, Ligong Han, Dimitris Metaxas
First, clone the repository and submodules. Then, set up a new conda environment and install all dependencies, as follows:
git clone --recursive https://github.com/statho/ScoreHMR.git
cd ScoreHMR
source install_environment.shDownload the pretrained model weights, and annotations for the datasets by running the following:
source download_data.sh
This will download all necessary data files, and place them in data/. Alternatively, you can download them from here and here. Besides these files, you also need to download the SMPL model. You will need the neutral model. Please go to the corresponding website and register to get access to the downloads section. Download the model, create a folder data/smpl, rename basicModel_neutral_lbs_10_207_0_v1.0.0.pkl to SMPL_NEUTRAL.pkl, and place it in data/smpl/.
Finally, if you wish to run the evaluation and/or training code, you will need to download the images/videos for the datasets. The instructions are mostly common with the description in here. We provide the annotations for all datasets, so you will only need to download the images/videos. Edit the IMG_DIR in score_hmr/configs/datasets.yml accordingly.
The following command will run ScoreHMR on top of HMR 2.0, using detected keypoints from ViTPose and bounding boxes from ViTDet, on all images in the specified --img_folder. For each image, it will save a rendering of all the reconstructed people together in the front view.
python demo_image.py \
--img_folder example_data/images \
--out_folder demo_out/imagesThe following command will first run tracking with 4D-Humans and 2D keypoint detection with ViTPose, and then run temporal model fitting with ScoreHMR on the video specified with --input_video. It will create a video rendering of the reconstructed people in the folder specified by --out_folder. It will also save intermediate results from 4D-Humans and ViTPose.
python demo_video.py \
--input_video example_data/videos/breakdancing.mp4 \
--out_folder demo_out/videosThe evaluation code is contained in eval/. We provide evaluation on 3 different settings with the following scripts:
eval_keypoint_fitting.pyis used in single-frame model fitting evaluation as in Tables 1, 2 & 5 of the paper.eval_multiview.pyis used to evaluate the multi-view refinement as in Table 3 of the paper.eval_video.pyis used to evaluate ScoreHMR in temporal model fitting as in Table 4 of the paper.
The evaluation code uses cached HMR 2.0 predictions, which can be downloaded from here or by running:
source download_hmr2_preds.shWe also provide example code for saving the HMR 2.0 predictions in the appropriate format in data_preprocessing/cache_hmr2_preds.py.
Evaluation code example:
python eval/eval_keypoint_fitting.py --dataset 3DPW-TEST --shuffle --use_default_ckptRunning the above command will compute the MPJPE and Reconstruction Error before and after single-frame model fitting with ScoreHMR on the test set of 3DPW.
The training code uses cached image features. First, extract the PARE image features for the training datasets:
python data_preprocessing/cache_pare_preds.py
Then, start training using the following command:
python train.py --name <name_of_experiment>
Checkpoints and logs will be saved to logs/.
Parts of the code are taken or adapted from the following repos:
If you find this code useful for your research, please consider citing the following paper:
@inproceedings{stathopoulos2024score,
title = {Score-Guided Diffusion for 3D Human Recovery},
author = {Stathopoulos, Anastasis and Han, Ligong and Metaxas, Dimitris},
booktitle = {CVPR},
year = {2024}
}
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent