lck1201 / seq2seq-3Dpose

My own Gluon reimplement of "Exploiting temporal information for 3D pose estimation"

Exploiting temporal information for 3D pose estimation

Reproduction of Exploiting temporal information for 3D pose estimation
Original implement is here

TODO:

• Provide trained model
• Refine project

Environment

python 3.7
mxnet-cu90 1.4.0
CUDA 9.0

Dependency

pip install pyyaml
pip install scipy
pip install matplotlib
pip install easydict

Dataset

1. Baidu Disk (code: kfsm) or Google Drive to download the HM3.6M annotation
2. Unzip data under data folder, and organize like this

${PROJECT_ROOT}
    `--data
        `--annot
            `--s_01_act_02_subact_01_ca_01
            `--s_01_act_02_subact_01_ca_02
            `-- ......
            `-- ......
            `-- ......
            `--s_11_act_16_subact_02_ca_04            

How-to-use

usage: train.py/test.py [-h] --gpu GPU --root ROOT --dataset DATASET [--model MODEL]
                        [--debug DEBUG]

optional arguments:
  -h, --help         show this help message and exit
  --gpu GPU          GPUs to use, e.g. 0,1,2,3
  --root ROOT        /path/to/project/root/
  --dataset DATASET  /path/to/your/dataset/root/
  --model MODEL      /path/to/your/model/, to specify only when test
  --debug DEBUG      debug mode

Train: python train.py --root /project-root

Test: python test.py --root /project-root --model /model-path

PS: You can modify default configurations in config.py.

Results

Since I don't have 2D pose estimate results on HM3.6M, I just experiment with 2D ground truth as input. My best result is 41.0mm , slightly higher than 39.2mm reported by paper

PS: LayerNorm is a component inside RNN cell. w/o=without