Exploring Collection of Sign Language Datasets: Privacy, Participation, and Model Performance

AnonySIGN: Novel Human Appearance Synthesis for Sign Language Video Anonymisation

“…Due to its visual medium, the collection of large-scale sign language corpora requires the storing of easily identifiable video data. Bragg et al suggests this introduces privacy concerns over data misuse, that significantly impacts the collection of sign language datasets [3].…”

“…Video effects of blackening [2] and pixelation [31] have been used to conceal sensitive information, which is unfeasible for extension to full data corpora. Bragg et al [3] suggest the use of greyscale or animoji filtering for video anonymisation, but neither provides full signer anonymity and both significantly impact sign comprehension. Focus groups with Deaf participants have suggested the use of actors or avatars to reproduce the data [38], but this requires labour-intensive work [20] and often results in non-realistic production [24].…”

“…As with the computational study of spoken languages, sign language research requires large-scale datasets to effectively tackle the tasks of Sign Language Translation (SLT) and Sign Language Production (SLP). However, due to privacy concerns regarding the visual nature of sign language datasets, there is a hesitancy to contribute to data collection [3].…”

“…Visual anonymisation of sign language datasets has been shown to increase the willingness of data participation due to the removal of signer-specific appearance [3]. However, previous approaches to sign language video anonymisation, such as pixelation [31], blackening [2] and greyscale filtering [3], have fallen woefully short of the mark. These methods significantly affect sign comprehension or worse, in the case of animoji filtering [3], deride the signers appearance.…”

“…However, previous approaches to sign language video anonymisation, such as pixelation [31], blackening [2] and greyscale filtering [3], have fallen woefully short of the mark. These methods significantly affect sign comprehension or worse, in the case of animoji filtering [3], deride the signers appearance.…”