I'm working on a final-year project involving NLP and text-to-image generation for American Sign Language (ASL). Most datasets I've found are video-based or limited to alphabet signs, but I specifically need a static image dataset containing complete ASL words rather than individual letters. Does a suitable dataset exist, or are there practical alternatives for building one?
2 Answers
A static dataset for complete ASL words may be difficult to find because many signs depend on hand movement, changes in hand shape, direction, and facial expression. Static images work reasonably well for fingerspelling and some simple signs, but they usually cannot capture the full meaning of a moving sign.
You may need to search university research archives and papers from sign-language or computer-vision labs, since some smaller datasets are not widely indexed. Another option is to extract representative keyframes from existing ASL video datasets and label them manually with a focused annotation tool. It would require substantial work and might lose movement information, but it could provide a usable prototype if a true static word-level dataset is unavailable.
That makes sense. It’s disappointing that a static dataset may not be suitable, but I appreciate the linguistic explanation.