From: A free geometry model-independent neural eye-gaze tracking system
Cameras | Lights | Calibration | Approach (mapping function) | Accuracy | Reference | Comments/Notes |
---|---|---|---|---|---|---|
1, 20 fps, 640×480 | 1 (+1) | screen divided in 2×4 zones | MI (2 GRNNs) | 5° H 8° V | [8]abc | Two concentric IR light rings are alternately turned on and off. Only one glint is used. |
1, 30 fps, 640×480 | 1 (0) | 12×16 grid | MI (1 MFNN) | 2.4° H 2.4° V | [17]bc | Special spectacles frame is needed; both the eyes are used. Accuracy is measured on testing points. |
1 low resol. | 1 | cursor moves | MI (2 MFNNs) | 1.5° | [16]abc | Accuracy is measured on testing points. |
1, 30 fps, 640×480 | 2 | 3×3 grid | MB | 0.9° | [9]ac | |
2 | 2 | MB | 0.68° | [9]ac | Preliminary simulations | |
1, 60 fps, 640×480 | 2-4 | 1 point | MB | ≈ 0.7° H ≈ 0.7° V ≈ 1° | [10]ac | Each calibration point produces a gaze estimation model. 17 one-point calibrations were performed. |
1, 60/120 fps, 640×480 | 4 (+1) | 5 points | NA | 0.5° | [22]ad | Tracks both eyes simultaneously. Camera and the IR sources are built in the monitor. |
1, 60/120 fps | 1 (+1) | 9 points | NA | 0.45°-0.70° | [23]ad | Typical and worst accuracy is reported. The POG estimation may require a dedicated computer. |
1, 25 fps, 640×480 | 3 | 4×5 grid | MI | ≤ 0.41° H ≤ 0.49° V ≤ 0.62° E | Proposed EGTSabc | Worst accuracy bounds measured on halfway 3×4 testing grid for 3 different geometric settings. |