Research Article
CNN-Based Pupil Center Detection for Wearable Gaze Estimation System
Table 1
Proposed CNN architectures.
| Name | Item | Conv1 | Conv2 | Conv3 | Conv4 | Conv5 | Full1 | Full2 | Out |
| Classification model A | Channel | 48 | 128 | 192 | 192 | 128 | 1024 | 1024 | 3 classes | Filter size | | | | | | — | — | — | Pooling size | | | | | | — | — | — | Normalization | yes | — | — | — | — | Yes | Yes | — | Dropout | — | — | — | — | — | Yes | Yes | — |
| Classification model B | Channel | 48 | 128 | 192 | 192 | 128 | 1024 | 1024 | 2 classes | Filter size | | | | | | — | — | — | Pooling size | | | | | | — | — | — | Normalization | yes | — | — | — | — | Yes | Yes | — | Dropout | — | — | — | — | — | Yes | Yes | — |
| Regression model | Channel | 96 | 256 | 512 | 512 | 512 | 4096 | 4096 | 2 reg. | Filter size | | | | | | — | — | — | Pooling size | | | — | — | | — | — | — | Normalization | yes | — | — | — | — | Yes | Yes | — | Dropout | — | — | — | — | — | Yes | Yes | — |
|
|