The project is seperated into a few different parts:
- Detector: Predict 3d poses in single camera videos for custom dataset creation
- Creator: RNN sequence generator
- Renderer: Blender project which animates a character with generated sequences
The project comes with a few sample sequences which already went through the process of 3D pose prediction. The following section will explain how the pool of sample sequences can be extended with custom video sequences.
We're using DVC. Please install the latest version.
The following directories are managed by DVC (look out for .dvc file):
data/train/video/example
data/train/video/deep-dance
Currently, there is only one remote configured, which works on a server with a locally shared folder between users:
/home/shared/dvc
If that folder does not exist, DVC can't pull the files.
The DVC workflow is very similar to Git. In a project folder, you could initialize and use DVC with
$ dvc init
$ dvc add data/*
$ dvc remote add -d local /home/shared/dvc
$ dvc run -d data -o model.p train.py
$ dvc push
$ git clone --recursive https://github.com/deep-dance/core.git
The repository reflects the parts mentioned above and seperates them into data, scripts, or other executables.
After cloning the respository, all data, managed by DVC also needs to be pulled.
$ dvc pull
Note, that the Docker container only works for a setup with nvidia graphic cards that can run cuda. See Docker instructions.
Make sure the folder VideoPose3D was cloned properly and download the pretrained model for 3D pose estimation.
cd VideoPose3D
mkdir checkpoint && cd checkpoint
wget https://dl.fbaipublicfiles.com/video-pose-3d/pretrained_h36m_detectron_coco.bin
for trajectory tracing also download addidional models here This link was posted by one of the maintainers of VideoPose3D in the issue comments here
This assumes the data structure to be
data/train/video/deep-dance/<dancer>/<tag-list>/input.mp4
data/train/video/deep-dance/<dancer>/<tag-list>/metadata.json
data/train/video/deep-dance/<dancer>/<tag-list>/pose2d.npz
data/train/video/deep-dance/<dancer>/<tag-list>/pose3d.mp4
It is recommended to prepare videos to match the Human3,6m data set as closely as possible. Video sequences should be saved in the folder mentioned above and should have a length of 30 seconds.
It's required to setup a motion database that reflects the dataset and contains metadata on every video in it. The current workflow requires the motion database to include the following metadata: name of the dancer and mapping of a video name to a list of tags. Both will be used while creating the dataset from custom custom videos. The motion database needs to be stored in a JSON file:
{
"dancers": [
{
"name": "adancer",
"videos": {
"V0001": [ "normal", "spot", "frontal", "standing" ]
}
}
]
There's a script that is used to preproces and store raw video files into the format explained above and it can be called like this:
python3 data/preprocess_videos.py --input-path videos/adancer --output-path train/video/deep-dance/adancer --database motion-db.json --metadata
The script takes an input path from which all video file are being picked up, scaled down and saved together with their metadata file. The folder structure that's being created depends on the tags given by the motion database. The metadata file contains a path to the original video it was created from and the tags given by the motion database.
Cut videos into sequences of 30 seconds, e.g.
ffmpeg -i input_raw.mp4 -ss 3:55 -to 4:25 -c copy data/train/seq_[index]/input_seq.mp4
It is desired to crop input videos such that they match Human3.6m in terms of the ratio between movement area and background. It still needs to be researched, if this is really necessary, or if the pose estimation results are still good enough without this step. Reducing the size of the input video has a positive effect on computation time in any case.
ffmpeg -i input_seq.mp4 -filter:v "crop=out_w:out_h:x:y" data/train/seq_[index]/input.mp4
out_w is the width of the output rectangle
out_h is the height of the output rectangle
x and y specify the top left corner of the output rectangle
Resize videos:
ffmpeg -i input_seq.mp4 -vf scale=800:600 output.avi
Before running 3D pose estimation, 2D poses need to be estimated. The following command will run the detectron2 predictor on all video files named input.mp4 in data/train/seq_[index] and create the numpy archive pose2d.npz, which contains detected boxes and keypoints. It does not overwrite exising pose2D.npz files, unless its told so by passing --overwrite.
Note:
By default the COCO-Keypoints/keypoint_rcnn_X_101_32x8d_FPN_3x model is used inside the detectron2 layer. Please refer to their documentation for other pretrained models and change paths in detector/predict_2d_pose.py accordingly.
python predict_2d_pose.py
python predict_3d_pose.py
The workflow is using DVC pipelines. All stages are included in dvc.yaml. Note: everything needs to happen in the Tensorflow Docker container.
The stage prepare was added issuing this:
dvc run -n prepare \
-p prepare.split,prepare.look_back \
-d src/prepare.py \
-o ../data/train/prepared \
python src/prepare.py
The stage train was added issuing this:
dvc run -n train \
-p train.data,train.epochs,train.batch_size,train.look_back,train.lstm_layer,train.mdn_layer,train.custom_loss \
-d src/train.py \
-M ../data/metrics/deep-dance-scores.json \
--plots-no-cache ../data/metrics/deep-dance-loss.json \
-o ../data/models \
-o ../data/metrics \
python src/train.py
The stage generate was added issuing this:
dvc run -n generate \
-p generate.seed,generate.steps_limit,generate.random_state,generate.look_back,generate.custom_loss \
-d src/generate.py \
-o ../data/generated \
python src/generate.py
There's no need to use dvc add for DVC to track stage outputs (../data/train/prepared/models in this case); dvc run already took care of this. You only need to run dvc push if you want to save them to remote storage, (usually along with git commit to version dvc.yaml itself).
dvc run exp needs a working Git config (because it could also push results to a Git remote directly). Therefor it is necessary to set a username and an email:
git config --global user.name "username"
git config --global user.email "username@mail.me"
cannot stash changes - there is nothing to stash.
One small workaround I have for this at the moment is make a small change to any file that is under git version control (but not related to any of the stages,etc). See issue (here)[iterative/dvc#5684].
python3.6 render_pose3d_matplot.py ../data/train/seq_001/pose3d.npz --frames 90
Several experiments with Blender have been done without any ground-breaking results. A component which translates keypoints into skeletal bone rotations for Blender needs to developed.
@erak @nikozoe