| Presentation | 2021-10-22 A scanpath prediction model using deep learning considering the context of the gazing objects Yuhei Ohsawa, Takeshi Kohama, |
|---|---|
| PDF Download Page |  PDF download Page Link |
| Abstract(in Japanese) | (See Japanese page) |
| Abstract(in English) | Since the human gaze is a biological signal which reflects internal states such as consciousness and attention, it is possible to develop technology for estimating human psychology by a gaze predicting system. In this study, we built a deep learning model that can reproduce human eye movement based on a constructive approach. First, we verified whether the encoder part of a typical image classification model could extract features that determine human eye movements from image statistics using a previous learning model. Next, we developed a deep learning model with a Seq2Seq structure that combines the encoder part of a typical image classification model with an RNN mechanism to construct a time series generation model with the context among objects and reproduced eye movements for still images. In order to evaluate the model performance, we calculated the cross-correlation coefficient between cumulative eye gaze distributions generated from the actual eye movements and the model outputs, and the average value for all validation data was 0.459. Since a small but positive correlation was found and temporal eye transitions were observed for objects and features that are likely to attract attention in the image, the proposed model can at least link image statistics with temporal characteristics of eye movements. |
| Keyword(in Japanese) | (See Japanese page) |
| Keyword(in English) | Visual system / Deep learning model / Gaze prediction / Scanpath |
| Paper # | HIP2021-43 |
| Date of Issue | 2021-10-14 (HIP) |
| Conference Information | |
| Committee | HIP |
|---|---|
| Conference Date | 2021/10/21(2days) |
| Place (in Japanese) | (See Japanese page) |
| Place (in English) | Online |
| Topics (in Japanese) | (See Japanese page) |
| Topics (in English) | |
| Chair | Shuichi Sakamoto(Tohoku Univ.) |
| Vice Chair | Yuji Wada(Ritsumeikan Univ.) / Sachiko Kiyokawa(Nagoya Univ.) |
| Secretary | Yuji Wada(NTT) / Sachiko Kiyokawa(NICT) |
| Assistant | Yuki Yamada(Kyushu Univ.) / Daisuke Tanaka(Tottori Univ.) / Ippei Negishi(Kanazawa Inst. of Tech.) |
| Paper Information | |
| Registration To | Technical Committee on Human Information Processing |
|---|---|
| Language | JPN |
| Title (in Japanese) | (See Japanese page) |
| Sub Title (in Japanese) | (See Japanese page) |
| Title (in English) | A scanpath prediction model using deep learning considering the context of the gazing objects |
| Sub Title (in English) | |
| Keyword(1) | Visual system |
| Keyword(2) | Deep learning model |
| Keyword(3) | Gaze prediction |
| Keyword(4) | Scanpath |
| 1st Author's Name | Yuhei Ohsawa |
| 1st Author's Affiliation | Kindai University(Kindai Univ.) |
| 2nd Author's Name | Takeshi Kohama |
| 2nd Author's Affiliation | Kindai University(Kindai Univ.) |
| Date | 2021-10-22 |
| Paper # | HIP2021-43 |
| Volume (vol) | vol.121 |
| Number (no) | HIP-211 |
| Page | pp.pp.69-74(HIP), |
| #Pages | 6 |
| Date of Issue | 2021-10-14 (HIP) |