Anonymous repository to host code and data to run the 3D hand gestures classification pipeline, based on Vispy and a ResNet-50 trained with Fast.ai, on the SFINGE 3D dataset.

• tested on Ubuntu 18.04 with CUDA already installed - inference performed on CPU (slower, but the setup is easier)

#> cd /tmp

Clone the SFINGE 3D dataset

#> git clone [email protected]:SFINGE3D/DatasetV1.git

Create a new virtualenv for this repo

#> mkdir /tmp/dynamic-hand-gestures-venv

#> python3 -m venv /tmp/dynamic-hand-gestures-venv

Activate the virtualenv

#> source /tmp/dynamic-hand-gestures-venv/bin/activate

Clone this repo

#> git clone [email protected]:dynamic-hand-gestures-classification/dynamic-hand-gestures-classification.git

Upgrade pip

#> pip install --upgrade pip

#> cd /tmp/dynamic-hand-gestures-classification/

Install the requirements

#> pip install -r requirements.txt

#> cd /tmp/dynamic-hand-gestures-classification/utilities/

#> ./conversion.py --filename /tmp/DatasetV1/Sequences/3.txt --csv-separator=';',

If you get CUDA errors here, such as:

ImportError: libcudart.so.9.0: cannot open shared object file: No such file or directory

it could mean that your CUDA drivers are too old for Pytorch 1.4.0, so downgrade it with:

#> pip3 install torch===1.2.0 torchvision===0.4.0 -f https://download.pytorch.org/whl/torch_stable.html

Unset LD_LIBRARY_PATH environment variable, just in case...

#> cd /tmp/dynamic-hand-gestures-classification

#> unset LD_LIBRARY_PATH

#> ./dynamic-hand-gestures.py ./utilities/unknown-3.csv.xz --dataset-path ./utilities/ --do-inference --model-name models/resnet-50-img_size-540-960-4a-2020-04-21_15.47.18-SFINGE3D-dataset-transfer-learning-from-our-dataset-data-augmentation-with-partial-gestures-and-noise.pkl --cuda-device cpu --inference-every-n-frames 20 --data-scale-factor 600.0 --data-x-offset -100 --data-y-offset -1250.0 --data-z-offset -250 --fps 10 --save-image-only-when-prob-greater-than 0.98

The VisPy visualizer should now go full screen and show the sequence as it is drawn and sent to the inference module (the ResNet-50 trained with Fast.ai).