ASL Fingerspelling Images (RGB & Depth)

Context

I've personally used part of this dataset to train a CNN to recognize ASL (American Sign Language) handshapes (github.com/MrGeislinger/ASLTransalation). Though the data have "incorrect" handshapes like the "G" handshape and produced by only 5 different individuals, the dataset is large enough to do a proof-of-concept of classifying ASL handshape images.

Content

The dataset contains cropped RGB images and depth data (collected from a Microsoft Kinect) of ASL (American Sign Language) handshapes corresponding to 24 letters of the English alphabet (note that "X" and "Z" are excluded since they rely on movement).

Note that this dataset was produced from 5 different non-native signers. Some of the handshapes are different from what traditional/native signers sign (e.g. "G" is frequently positioned differently from traditional signers). Data were retrieved from Nicolas Pugeault's website, referenced below.

Acknowledgements

Original paper: Pugeault, N., and Bowden, R. (2011). Spelling It Out: Real-Time ASL Fingerspelling Recognition In Proceedings of the 1st IEEE Workshop on Consumer Depth Cameras for Computer Vision, jointly with ICCV'2011.

"ASL Finger Spelling Dataset"

Inspiration

My hope is to see more work done in ASL and other sign languages recognition. It'd be great to one day see a tool that can be versatile for sign languages as voice-to-text technology is today.

데이터와 리소스