Hello everyone! I urgently need guidance for my final-year project involving NLP and text-to-image generation for American Sign Language (ASL). Most datasets I have found contain videos or focus only on fingerspelling and alphabet signs. I specifically need a word-level ASL dataset made up of static images rather than video clips. Does anyone know of a suitable dataset, research archive, or practical way to create one from existing resources?
2 Answers
A static image dataset may not represent most ASL words accurately. In sign languages, meaning often depends on hand movement, changes in hand shape, position, and facial expression. Alphabet and number signs can usually be classified from a single frame because their shapes remain relatively stable, but many word-level signs require motion and are better represented as video or sequences of frames.
You may need to search university research repositories and specialized computer-vision or language labs, since smaller or older datasets are sometimes mentioned in papers without being easy to find. Another option is to extract carefully selected keyframes from existing ASL video datasets and annotate them, although this would require substantial manual work and may lose important movement information. A sequence-based approach using short clips or ordered keyframes could be a more realistic compromise than treating each word as one isolated image.
Thanks for explaining that. It is disappointing that a static dataset may not be linguistically sufficient, but I understand why movement is essential for many signs.