Automatic recognition of vehicle
attributes using machine learning
Mgr. Jan Sedlák
System
Vstupní data Předzpracování
Klasifikace
Uložení a prezentace
Extrakce informace
Input data
Kamera
SERVER
Kamera
Region of interest
 Pixel intensity values
128
54
Preprocessing
How to classify?
 Let’s go deep…
Deep learning
 Test some recommended configurations
 hidden = [200,200]
 hidden = [512]
 hidden = [64,64,64]
 hidden = [32,32,32,32,32]
 hidden = [1024,512,256]
Overfitting and dropout
Use dropout ratio
Use adaptive rate
What about results?
Random forest?
Deep learning - features
Advantages and disadvantages
Bad MSE Good MSE
Detecting anomalies
Reconstruction example
Reconstruction error
Detected anomalies