Disha Ai Terrain Segmentation

Semantic Segmentation for Offroad Environments

This model segments offroad terrain images into 10 classes:

  • Background, Trees, Lush Bushes, Dry Grass, Dry Bushes
  • Ground Clutter, Logs, Rocks, Landscape, Sky

Architecture:

  • Backbone: DINOv2-ViT-Small (frozen)
  • Head: Mask2Former with dual-resolution decoder
  • Resolution: 672×392 pixels
  • Parameters: ~15M trainable

Technical Details:

  • Training: Inverse-frequency log weighting + Focal loss + Dice loss + Boundary loss
  • Data Augmentation: Copy-paste for rare classes, color jitter, rotation, crop
  • Inference: Efficient Mask2Former with low-res attention and high-res masks

How to use:

  1. Upload an offroad/outdoor image
  2. Click "Segment" or use the example images
  3. View the colored segmentation mask and overlay

Model trained on custom offroad dataset with class-balanced loss.