yogikushle/aisegmentation

Offroad Terrain Segmentation Model

This is a semantic segmentation model trained to identify offroad terrains from images.

Model Details

• Model Architecture: DeepLabV3 with ResNet50 backbone
• Pre-training: Initialized with weights pre-trained on COCO dataset
• Dataset: 'Offroad_Segmentation_Training_Dataset'
• Input: RGB images (540x960 pixels)
• Output: Segmentation mask with 2 classes (e.g., background, offroad terrain)
• Training Epochs: 10
• Batch Size: 16
• Learning Rate: 0.001

Training Metrics

• Final Training Loss: 0.0682
• Final Validation Loss: 0.0785
• Final Training Mean IoU: 0.2789
• Final Validation Mean IoU: 0.2690

How to use

import torch
import torchvision.models.segmentation as models
from torchvision import transforms
from PIL import Image
import cv2
import numpy as np

# Load the model architecture
model = models.deeplabv3_resnet50(pretrained=False) # Set pretrained=False as we load custom weights
model.classifier[4] = torch.nn.Conv2d(256, 2, kernel_size=1) # Adjust output channels for 2 classes

# Load the state dictionary
model_path = "deeplabv3_resnet50_offroad.pth" # Path to your saved model
model.load_state_dict(torch.load(model_path, map_location=torch.device('cpu')))
model.eval()

# Preprocessing transformations
transform = transforms.Compose([
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])

def predict_mask(image_path, model, transform):
    image = Image.open(image_path).convert("RGB")
    original_size = image.size

    # Resize to model's expected input size (960x540 for this model, or handle dynamic resizing)
    # For simplicity, assuming model input was trained on fixed size, let's resize
    image_tensor = transform(image).unsqueeze(0) # Add batch dimension

    with torch.no_grad():
        output = model(image_tensor)['out']

    # Get the predicted class for each pixel
    predicted_mask = torch.argmax(output.squeeze(), dim=0).cpu().numpy()

    # Resize mask back to original image size if necessary
    predicted_mask_resized = cv2.resize(predicted_mask.astype(np.uint8), original_size, interpolation=cv2.INTER_NEAREST)

    return predicted_mask_resized

# Example usage:
# Assuming you have an image 'test_image.jpg'
# mask = predict_mask('test_image.jpg', model, transform)
# plt.imshow(mask)
# plt.show()