HDNet_TikTok

This repository is the official tensorflow python implementation of "Learning High Fidelity Depths of Dressed Humans by Watching Social Media Dance Videos" in CVPR 2021 (Oral Presentation) (Best Paper Nominated).

TikTok1

This codebase provides:

  • Inference code
  • Training code
  • Visualization code

Requirements

(This code is tested with tensorflow-gpu 1.14.0, Python 3.7.4, CUDA 10 (version 10.0.130) and cuDNN 7 (version 7.4.2).)

  • numpy
  • imageio
  • matplotlib
  • scikit-image
  • scipy==1.1.0
  • tensorflow-gpu==1.14.0
  • gast==0.2.2
  • Pillow

Installation

Run the following code to install all pip packages:

pip install -r requirements.txt 

In case there is a problem, you can use the following tensorflow docker container "(tensorflow:19.02-py3)":

sudo docker run --gpus all -it --rm -v local_dir:container_dir nvcr.io/nvidia/tensorflow:19.02-py3

Then install the requirements:

pip install -r requirements.txt 

Inference Demo

Input:

The test data dimension should be: 256x256. For any test data you should have 3 .png files: (For an example please take a look at the demo data in "test_data" folder.)

  • name_img.png : The 256x256x3 test image
  • name_mask.png : The 256x256 corresponding mask. You can use any off-the-shelf tools such as removebg to remove the background and get the mask.
  • name_dp.png : The 256x256x3 corresponding DensePose.

Output:

Running the demo generates the following:

  • name.txt : The 256x256 predicted depth
  • name_mesh.obj : The reconstructed mesh. You can use any off-the-shelf tools such as MeshLab to visualize the mesh. Visualization for demo data from different views:

mesh2

  • name_normal_1.txt, name_normal_2.txt, name_normal_3.txt : Three 256x256 predicted normal. If you concatenate them in the third axis it will give you the 256x256x3 normal map.
  • name_results.png : visualization of predicted depth heatmap and the predicted normal map. Visualization for demo data:

0043_results

Run the demo:

Download the weights from here and extract in the main repository or run this in the main repository:

wget --load-cookies /tmp/cookies.txt "https://docs.google.com/uc?export=download&confirm=$(wget --quiet --save-cookies /tmp/cookies.txt --keep-session-cookies --no-check-certificate 'https://docs.google.com/uc?export=download&id=1UOHkmwcWpwt9r11VzOCa_CVamwHVaobV' -O- | sed -rn 's/.*confirm=([0-9A-Za-z_]+).*/\1\n/p')&id=1UOHkmwcWpwt9r11VzOCa_CVamwHVaobV" -O model.zip && rm -rf /tmp/cookies.txt

unzip model.zip

Run the following python code:

python HDNet_Inference.py

From line 26 to 29 under "test path and outpath" you can choose the input directory (default: './test_data'), ouput directory (default: './test_data/infer_out') and if you want to save the visualization (default: True).

More Results

TikTok2

Training

To train the network, go to training folder and read the README file

MATLAB Visualization

If you want to generate visualizations similar to those on the website, go to MATLAB_Visualization folder and run

make_video.m

From lines 7 to 14, you can choose the test folder (default: test_data) and the image name to process (default: 0043). This will generate a video of the prediction from different views (default: "test_data/infer_out/video/0043/video.avi") This process will take around 2 minutes to generate 164 angles.

Note that this visualization will always generate a 672 × 512 video, You may want to resize your video accordingly for your own tested data.

Citation

If you find the code or our dataset useful in your research, please consider citing the paper.

@InProceedings{Jafarian_2021_CVPR_TikTok,
    author    = {Jafarian, Yasamin and Park, Hyun Soo},
    title     = {Learning High Fidelity Depths of Dressed Humans by Watching Social Media Dance Videos},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2021},
    pages     = {12753-12762}} 

GitHub

https://github.com/yasaminjafarian/HDNet_TikTok