2023 Fiscal Year Research-status Report
Developing an integrated account of intentions and affordances for a model of visual attention
| Project/Area Number |
23K11169
|
|---|
| Research Institution | Okayama University |
|---|
Principal Investigator |
Yucel Zeynep 岡山大学, 環境生命自然科学学域, 准教授 (20586250)
|
|---|
| Project Period (FY) |
2023-04-01 – 2025-03-31
|
| Keywords | saliency / affordance / segmentation |
|---|
| Outline of Annual Research Achievements |
We examined the performance of four recent saliency models EML-NET, SalGAN, DeepGaze IIE, and DeepGaze on images of hand tools. These objects have distinct segments with various roles, and studies suggest that tool segments inherently attract human attention. We tested the models on a dataset containing both tool and non-tool images, then compared their predictions with human gaze data using six criteria. The results show that the models often struggle to predict saliency accurately for tool images compared to non-tool images. This suggests a need to address this limitation in saliency modeling for tool-specific contexts.
|
| Current Status of Research Progress |
Current Status of Research Progress
2: Research has progressed on the whole more than it was originally planned.
Reason
We have demonstarted that the existing state-of-the-art saliency models are not as efficient in representing eye gaze patterns over tool images as they are in representing the eye gaze patterns over other images from ordinary daily life scenes. This justifies an effort to improve the existing methods.
|
| Strategy for Future Research Activity |
We focus on the state-of-the-art visual saliency prediction model of DeepGaze IIE and make an effort to refine it to account for this bias. Since the integration of transfer learning into saliency prediction over the last decade has notably enhanced prediction performance, we will initially curate a custom image data set featuring tools, non-tools and ambiguous images and record empirical gaze data from human participants to be used in fine-tuning. In this way, we will improve the model’s performance for this specific stimulus category and evaluate it by IG and NSS metrics.
|
