Figure 1.
The architecture of GazeTR
Figure 2.
The detailed architecture of AG-HybridNet.
Figure 3.
The detailed architecture of the CNN branch.
Figure 4.
Schematic diagram of the Reparametrized Partial Convolution (RPConv) structure.
Figure 5.
Network architecture diagram showing the RPConv and TDConv Block structure.
Figure 6.
The detailed architecture of the attention mechanism.
Figure 7.
MPIIFaceGaze and Gaze360 Loss Convergence Curves.
Figure 8.
Comparison of Mean Angular Error on the MPIIFaceGaze Dataset.
TABLE I.
Comparative Experimental Results on the MPIIFaceGaze Datase
| Model | Mean Angular Error (°) |
|---|---|
| MPIIGaze | 5.40 |
| Dilated-Net | 4.80 |
| CA-Net | 4.10 |
| AGE-Net | 4.09 |
| GazeTR | 4.00 |
| L2CS-Net | 3.92 |
| Res-Swin-Ge | 3.75 |
| Ours | 3.72 |
TABLE II.
Comparative Experimental Results on the Gaze360 Dataset
| Model | Mean Angular Error (°) |
|---|---|
| Full-Face | 14.99 |
| Dilated-Net | 13.73 |
| RT-Gene | 12.26 |
| Gaze360 | 11.40 |
| Bot2L-Net | 11.53 |
| Ours | 10.82 |
TABLE III.
Comparison of Parameters and FLOPS for Different Models
| Model | Mean Angular Error (°) | Parameters | FLOPs |
|---|---|---|---|
| Dilated-Net | 4.80 | 3.920 | 3.153 |
| GazeTR | 4.00 | 11.394 | 1.834 |
| Ours | 3.72 | 21.201 | 1.505 |
Figure 9.
Comparison of Mean Angular Error and FLOPs