-
AlphaDAPR: An AI-based Explainable Expert Support System for Art Therapy논문 2024. 4. 1. 09:39
https://dl.acm.org/doi/abs/10.1145/3581641.3584087
I. Introduction
Sketch plays an interpreting role in understanding an individual's psychological and cognitive state. Drawing can reflect preconscious or unconscious material. Experts seek to identify psychological indicators based on the predefined scoring scales that consider how a participant expresses a human figure and its environment in a sketch. However, this is time-consuming. Thus, an automatic analysis system is in need. Three functions should be included in the system. First, a drawing should be accurately analyzed. Second, the score that reflects the psychological state should be automatically calculated. Third, the final analysis results should be provided.
II. Data
1. Curating publicly available sketches. (from six papers)
2. Creating new sketches by recruiting participants.
3. Augmentation.
- augmentation was conducted in a way that the drawn figure is substituted by the existing data below
* QuickDraw dataset
7,500 images of rain, umbrella, puddle, lightning, and cloud.
https://quickdraw.withgoogle.com/data
* TU-Berlin dataset
Eitz, M., Hays, J., & Alexa, M. (2012). How do humans sketch objects? ACM Transactions on graphics (TOG), 31(4), 1-10.
20,000 unique sketches evenly distributed over 250 object categories.
https://dl.acm.org/doi/abs/10.1145/2185520.2185540
* Rakhmanov, O., Agwu, N. N., & Adeshina, S. (2020, May). Experimentation on hand drawn sketches by children to classify Draw-a-Person test images in psychology. In The Thirty-Third International Flairs Conference.
Gathered 1,000 Draw-a-Person test images, but released images not found.
III. Model
Yolo-v5 achieved the best score.
IV. Evaluation
Demographics
Evaluation Aggregation
Impression I got
Half of the participants did not recognize score-related information as useful. Rather, they recognized supplementary information such as participant information, sketch replay, and the number and average length of lines as useful. On the one hand, it means that participants wanted to make decisions themselves, as this information did not make any decision itself. On the other hand, it means that they did not trust score-related information. Trust depends on accuracy. The mean average precision of the best model, Yolo-v5, was 50.46. Of course, this is not satisfying. We also need to consider the fact that the AI model cannot take the background of the participant into account like art therapists do when they counsel clients and interpret drawings. Then, can we make a model that takes both the drawing and the background into account at the same time? It's possible to put the images and the texts in the model at the same time, but we still do not know if the model can give us useful information compared to human insight.
'논문' 카테고리의 다른 글