Poses with backwards bent legs

[da2r-20]

Hi, I ran SHAPY on my data

I have 3 issues that occur frequently on most of my test samples:

1. The head pose frequently leans back
2. The knees are frequently bent backwards.
3. Hands pose are not accurate

I've seen this happening slightly on the demo set but not as frequent as on my data.

I was wondering if I'm missing anything.

Thanks in advance!

SMPLify-X vs SHAPY compared pose examples

Configuration

Input data (images and keypoints with SHAPY and SMPLify-X results): data.zip

genders.yaml

...
"woman-01": "female"
"woman-02": "female"
...

Run command:

python demo.py \
    --save-vis true \
    --save-params true \
    --save-mesh true \
    --split test \
    --datasets openpose \
    --output-folder ../samples/public-testset/shapy_output \
    --exp-cfg configs/b2a_expose_hrnet_demo.yaml \
    --exp-opts \
            output_folder=../data/trained_models/shapy/SHAPY_A \
            part_key=pose \
            datasets.pose.openpose.data_folder=../samples \
            datasets.pose.openpose.img_folder=public-testset/images \
            datasets.pose.openpose.keyp_folder=public-testset/keypoints \
            datasets.batch_size=1 \
            datasets.pose_shape_ratio=1.0

[muelea]

SHAPY focuses on body shape, so the examples you shared look like what I would expect. If you wanted better hand poses and facial expressions better use PXIE.

[datased2019]

I am also give a try on my own dataset and find out that the measurement is not accurate especially height, legs and etc. Any comment on that? Thx.

[muelea]

More information would be helpful here.. First, what is not accurate the regressor, the virtual measurements or language to shape? And do you have examples? What accuracy would you expect?