Awesome

CLIFF [ECCV 2022 Oral]

Introduction

This repo is highly built on the official CLIFF and contains an inference demo, and further adds accurate detector and multi-person tracker. For post-processing, motion interpolation and smooth are supported for better visualization results.

CLIFF: Carrying Location Information in Full Frames into Human Pose and Shape Estimation.

Major features

• [08/20/22] Support YOLOX as single-person detector, better performance on single frame.
• [08/20/22] Support ByteTrack as multi-person tracker, better performance on person Re-ID.
• [08/20/22] Support linear interpolation as motion completion method, especially for occlusion.
• [08/20/22] Support Smooth-Net as post-processing motion smooth for decresing motion jittering.
• [09/29/22] Support SMPLify fitting given GT/Pred 2D keypoints for improving the quality of estimated SMPL params.
• [01/31/23] Further support motion smooth for SMPL pose and translation besides of 3D joints.

Preparation

conda create -n cliff python=3.10
pip install -r requirements.txt

1. Download the SMPL models for rendering the reconstructed meshes
2. Download the pretrained checkpoints to run the demo [Google Drive]
3. Install MMDetection and download the pretrained checkpoints
4. Install MMTracking and download the pretrained checkpoints

Finally put these data following the directory structure as below:

${ROOT}
|-- data
    smpl_mean_params.npz
    |-- ckpt
        |-- hr48-PA43.0_MJE69.0_MVE81.2_3dpw.pt
        |-- res50-PA45.7_MJE72.0_MVE85.3_3dpw.pt
        |-- hr48-PA53.7_MJE91.4_MVE110.0_agora_val.pt
    |-- smpl
        |-- SMPL_FEMALE.pkl
        |-- SMPL_MALE.pkl
        |-- SMPL_NEUTRAL.pkl
|-- mmdetection
    |-- checkpoints
        |-- yolox_x_8x8_300e_coco_20211126_140254-1ef88d67.pth
|-- mmtracking
    |-- checkpoints
        |-- bytetrack_yolox_x_crowdhuman_mot17-private-half_20211218_205500-1985c9f0.pth

Demo

We provide demos for single-person and multi-person video.

Single-person

Run the following command to test CLIFF on a single-person video:

python demo.py --ckpt data/ckpt/hr48-PA43.0_MJE69.0_MVE81.2_3dpw.pt \
               --backbone hr48 \
               --input_path test_samples/01e222117f63f874010370037f551497ac_258.mp4 \
               --input_type video \
               --save_results \
               --make_video \
               --frame_rate 30

Multi-person

Use the --multi flag to support multi-person tracking, --infill flag to support motion infill, --smooth flag to support motion smooth. Run the following command to test CLIFF on a multi-person video with post-processing:

python demo.py --ckpt data/ckpt/hr48-PA43.0_MJE69.0_MVE81.2_3dpw.pt \
               --backbone hr48 \
               --input_path test_samples/62883594000000000102c16c.mp4 \
               --input_type video \
               --multi \
               --infill \
               --smooth \
               --save_results \
               --make_video \
               --frame_rate 30

SMPLify Fitting

As the same as SPIN, we apply SMPLify fitting after CLIFF, OpenPose format 2D Keypoints are required for convinence.

python3 demo_fit.py --img=examples/im1010.jpg \ 
                    --openpose=examples/im1010_openpose.json

Citing

@Inproceedings{li2022cliff,
  Title     = {CLIFF: Carrying Location Information in Full Frames into Human Pose and Shape Estimation},
  Author    = {Li, Zhihao and Liu, Jianzhuang and Zhang, Zhensong and Xu, Songcen and Yan, Youliang},
  Booktitle = {ECCV},
  Year      = {2022}
}

Contact

If you have problems about usage, feel free to open an issue or directly contact me via: haofanwang.ai@gmail.com. But please note that I'm NOT the author of CLIFF, so for any question about the paper, contact the author.