This repository explores CNN-based pupil center estimation from infrared eye images captured with a smartphone-based head tracker.
The project focuses on estimating pupil center coordinates from eye images and comparing how different image-enhancement techniques affect localization error.
The CNN architecture uses:
- Five convolutional layers with 3x3 filters and stride 1
- Average pooling with 2x2 windows and stride 2
- Batch normalization and dropout after each layer
- One fully connected layer with 2048 units
- Euclidean distance between true and predicted pupil centers as the loss function
The project compares several preprocessing approaches:
- Histogram equalization
- Power-law transformation
- Adaptive histogram equalization
- Adaptive histogram equalization combined with power-law transformation
Raw infrared eye image and histogram examples:
Adaptive histogram equalization:
Power-law transformation:
Histogram equalization:
Adaptive histogram equalization plus power-law transformation:
Mean pixel error after training with different enhancement techniques:
