Run on complete dataset

Training on the complete dataset shows some significant improvement in the results. The overall accuracy of the prediction has increased. We can also observe, some of the segmentation of the objects are true to the shape of the object, like cars. We can also notice the smaller objects like street lamps are more clearly recognized.

Add Tversky loss to the Cross Entropy Loss

Tversky Loss is a generalization of DICE Loss and it gives the information of how well a class was predicted in a batch. Cross Entropy loss tells how well each pixel was classified. Using both the loss functions together will help the model understand from both perspectives on what is being wrongly classified. The results can be seen as per this. Classes like lanes and better identified by the model and it has also started identifying things like fencing on the side of the road.

Create a custom loss function

Made a new loss function based on the cross-entropy loss and DICE coefficient. The class weights of the cross-entropy loss for each class is updated for each batch using the DICE coefficient. This made the model start prioritizing the classes which are under-represented and it can be seen in the results. Although the accuracy fell by a small amount, the model is able to recognize a higher number of objects. For an autonomous driving car, I feel recognizing objects like lanes, street lights, and signs which are under-represented by the pixel count when compared to things like roads are important.

The depth of the model was changed back to 5 because having a depth of 7 made no change to the performance.

Increase depth to 7

Increasing the depth of the UNet did not make any significant change in the result. The only noticeable change was the increase in the model size from 124 MB to 1.9 GB.

UNet with batch normalization

Adding batch normalization is helping the model converge faster and also give a significant improvement in the accuracy.

Baseline model

A baseline model using a UNet for segmentation with a depth of 5. Most of the large objects have been identified well but the smaller objects are not.