Multiclass Semantic Segmentation for Road Surface Detection

1. Identified road surfaces and 13 different classes like speed bumps, paved, unpaved, markings, water puddles, potholes, etc.
2. Trained the model for semantic segmentation on Unet architecture along with backbone architectures like Resnet, InceptionNet and VGGnet.
3. Added mask to images to show the classes according to their respective colors.

Model Training

Sl. No.	Model	Epochs	Mean IoU Score on CV
1.	UNet	20	0.26527
2.	UNet with ResNet18	10	0.6309
3.	UNet with ResNet34	100	0.7297
4.	UNet with InceptionNetV3	20	0.6633
5.	UNet with VGGnet16	20	0.6604

Markings

Sl. No.	Color	Category
1.	Black	Background
2.	Light Blue	Road Asphalt
3.	Greenish Blue	Paved Road
4.	Peach/Light Orange	Unpaved Road
5.	White	Road Marking
6.	Pink	Speed Bump
7.	Yellow	Cats Eye
8.	Purple	Storm Drain
9.	Cyan	Manhole Cover
10.	Dark Blue	Patches
11.	Dark Red	Water Puddle
12.	Red	Pothole
13.	Orange	Cracks

Deployed On Hugging Face: Link