Shrec 2021 submission

A. D’Eusanio, A. Simoni, S. Pini, G. Borghi, R. Vezzani, R. Cucchiara

We propose a transformer-based architecture for the dynamic hand gesture recognition task, using the features provided by the Leap Motion camera.

General script and test parameters:

python main.py \
--hypes /path/to/test_realtime.json \
--phase test_realtime \
--resume /path/to/checkpoints.pth

We propose 3 different settings, you can test them with the following commands:

• python main.py --hypes ./hyperparameters/shrec/test_realtime_a.json
• python main.py --hypes ./hyperparameters/shrec/test_realtime_b.json
• python main.py --hypes ./hyperparameters/shrec/test_realtime_c.json

Results are saved in the outputs folder.

Inside the checkpoints folder (download it from here) you can find the 3 files related to our 3 submissions:

• best_full_dataset.pth
• best_full_dataset_norm.pth
• best_full_dataset_moredata.pth

Additional information about the config file:

• "n_features": [20, 13] -> if resuming submission a (default normalization) or b (normalization with hand dimension)
• "n_features": [20, 32] -> if resuming submission c (default normalization + joints distances)
• "norm_hand" -> true if resuming submission b (hand dimension normalization) else false
• "distances" -> true if resuming submission c (joint distances as additional features) else false