Emergency Vehicle Classification

The competition was hosted on Analytics Vidhya. The task was to classify vehicles into emergency and non-emergency categories.

Team Name : The Alphas

Authors : Ashay Ajbani and Pritam Rao

Competition Type : Image Classification

Framework : Keras

Solution Type : Model Ensembling

Number of Classes : 2

Pretrained Models Used : VGG16 and NASnet Large

Private Leaderboard Rank : 36/10000

Our Approach

We ensembled 5 models to make final predictions.

Models :

• Model 1 : Fine Tuning VGG16
• Model 2 : Fine Tuning VGG16 with l2 regularization
• Model 3 : k-fold validation
• Model 4 : Fine Tuning NASnet
• Model 5 : Fine Tuning NASnet with l2 regularization

The predictions taken by ensembling above 5 models resulted in an accuracy greater than any of the individual models.

Techniques used to reduce overfitting :

• Data Augmentation
• BatchNormalization
• Dropout
• l2 Regularization

Accuracy :

• Model 1 : 0.9498
• Model 2 : 0.9546
• Model 3 : 0.9444
• Model 4 : 0.9548
• Model 5 : 0.9514
• After Ensembling (final) : 0.9688

Click here to access the code