ASLLVD-Skeleton-20
The ASLLVD-Skeleton-20 is a subset of the ASLLVD-Skeleton database and is composed of 20 selected signs detailed in the table below.
| Dataset | Selected signs |
|---|---|
| ASLLVD | adopt, again, all, awkward, baseball, behavior, can, chat, cheap, cheat, church, coat, conflict, court, deposit, depressed, doctor, don’t want, dress, enough |
The result of each step is similar to that obtained in creating the ASLLVD-Skeleton, but only for the above signs. For more details, click here.
Further details on the creation of this dataset can be found in the paper Spatial-Temporal Graph Convolutional Networks for Sign Language Recognition. If you have any questions, please feel free to contact me at cca5@cin.ufpe.br.
1. Acquire samples
Click here for more details on this step.
2. Segment samples
Taken from the ASLLVD dataset. See American Sign Language Lexicon Video Dataset (ASLLVD) and for direct download of the most recent ASLLVD data (with corrections and revisions since the original release in 2012), access https://dai.cs.rutgers.edu/dai/s/signbank.
Extremely important: The ASLLVD video data are subject to Terms of Use: http://www.bu.edu/asllrp/signbank-terms.pdf.
By downloading these video files, you are agreeing to respect these conditions. In particular, NO FURTHER REDISTRIBUTION OF THESE VIDEO FILES is allowed.
Click here to download the output for the 20 selected signs.
3. Estimate skeletons
Click here to download the output for the 20 selected signs.
4. Filter keypoints
Click here to download the output for the 20 selected signs.
5. Split dataset
In order to make the division more suitable for the few samples of the above 20 signals, the following ratio was applied in the division of this subset
- 67% for training;
- 33% for testing;
Click here to download the output of this step.
6. Normalize and serialize
Click here to download the output for the 20 selected signs.
If you encounter any problems downloading the above files, click here to explore the files directory.