CNN Architecture for Pupil Center Estimation in eye images extracted from head tracker of a smartphone

*5 Convolutional Layers with a stride of 1 and 3x3 filters

• 1 Fully Connected Layer with 2048 units
  
• Average Pooling of 2x2 with a stride of 2
  
• Batch Normalization and Dropout after every layer
  
• Loss Function – Euclidean Distance between true and predicted labels
  

Raw images of the pupil obtained from a infrared camera in a smartphone along with their histograms

Result of applying different image enhancement techniques such as Histogram Equalisation, Power Law, Adaptive Histogram Equalisation

• Adaptive Histogram Equalisation

* Power Law

* Histogram Equalisation

* Adaptive Histogram Equalisation plus Power Law

Mean Pixel Error after training on the same dataset but with different image enhancement techniques