Model description:

This model expects a cropped image of an eye as input.

It returns with the sclera, iris and pupil bounding box.

This model can be used for both colour and infrared images.

​

Business case:

Localisation of the iris, can be used to implement biometric identification systems.
This model also lets you measure pupil dilation, which can be used to improve emotion recognition or measure cognitive workload. 

​

Number of classes: 3 (sclera, iris, pupil)

 

Metrics:
 detections_count = 63837, unique_truth_count = 31416
class_id = 0, name = sclera, ap = 99.95%         (TP = 10102, FP = 4478)
class_id = 1, name = iris, ap = 95.31%           (TP = 10421, FP = 10445)
class_id = 2, name = pupil, ap = 60.28%          (TP = 10270, FP = 9892)
for conf_thresh = 0.25, precision = 0.55, recall = 0.98, F1-score = 0.71
for conf_thresh = 0.25, TP = 30793, FP = 24815, FN = 623, average IoU = 48.97 %
IoU threshold = 50 %, used Area-Under-Curve for each unique Recall
mean average precision (mAP@0.50) = 0.851784, or 85.18 %

​

System requirements:

Inference time using CPU: 300 ms (on HP Laptop 15-DA0042NH (Processor: Intel(R) Core(TM) i7-8550U CPU))