Automatic detection of the visual gaze components of joint attention in observational, naturalistic child language acquisition data
| Type of publication: | Conference paper |
| Citation: | Dickerman_BUCLD_2025 |
| Publication status: | Accepted |
| Booktitle: | Boston University Conference on Language Development |
| Year: | 2024 |
| Month: | November |
| Abstract: | This study aims to describe gaze behaviors in joint attention interactions within natural settings and assess the potential of machine learning tools for automated annotation of gazepoints and gaze behaviours. Previous work in the area has relied on eye-tracking equipment, limiting the ‘naturalness’ of the data. The authors compiled a large dataset from an observational, naturalistic language acquisition corpus collected in the homes of six children aged 2 to 4 years, manually annotating gaze components such as shared attention and eye contact. Frame-wise analysis of the data showed, for example, that gaze points are closer during joint attention, despite the messy, naturalistic dataset. This distinction was successfully captured by pre-trained gaze models. Results overall indicate that pre-trained gaze detection models perform relatively well on the new datatype, and that fine-tuning improves their effectiveness. This suggests promising potential for future work on naturalistic gaze behaviour in language acquisition. |
| Keywords: | |
| Projects |
Idiap AI4Autism |
| Authors | |
| Added by: | [UNK] |
| Total mark: | 0 |
|
Attachments
|
|
|
Notes
|
|
|
|
|