Visual Learning - Semantic Segmentation for Self-Driving Cars

Streamlit Application

In the code we have implementated two models from scratch

1. Unet
2. Deep lab V3
3. Deep lab V3 - ViT

Results

1. Unet

• Train Accuracy: 98.34
• Train Dice score: 0.83545
• Test Accuracy 97.24
• Test Dice score: 0.83259

2. Deep lab V3 model

• Train Accuracy: 88.35
• Train Dice score: 0.79730
• Test Accuracy: 88.39
• Test Dice score: 0.79736

3. Deep lab V3 - ViT model

• Train Accuracy: 83.97
• Train Dice score: 0.77760
• Test Accuracy: 84.07
• Test Dice score: 0.77762

Test (Left) and Train (right) loss for Deep Lab

Test (Left) and Train (right) loss for Unet

Test (Left) and Train (right) loss for Deep Lab VIT

Predictions on Deep Lab

Predictions on UNet

Predictions on Deep Lab V3 - ViT