🏜️ Desert Terrain Analysis

Desert Semantic
Segmentation

Every pixel classified. 10 terrain classes. Powered by DINOv2 vision transformer.

10 Classes
960×540 Resolution
DINOv2 Backbone
View Results See Metrics
Background
Trees
Lush Bushes
Dry Grass
Dry Bushes
Ground Clutter
Logs
Rocks
Landscape
Sky
Metrics

Training Dashboard

Model performance across training rounds with iterative improvements.

🎯
—
Mean IoU
Best
🔬
—
Mean Dice
Best
📐
—
Pixel Accuracy
Best
⏱️
—
Epochs Trained
Per Round

📈 Round-over-Round Improvement

Per-Class IoU — Round 1 vs Round 2

Loss Curves — Both Rounds

📊 Round 1 — Training Plots

Round 1 training curves

📊 Round 2 — Training Plots

Round 2 training curves

📊 Round 1 — Per-Class IoU

Round 1 per-class IoU

📊 Round 2 — Per-Class IoU

Round 2 per-class IoU

Class Performance Breakdown (Best Round)

Analysis

Failure Case Analysis

Understanding where and why the model struggles — key to iterative improvement.

About

About the Project

Pixel-level understanding of desert terrain using deep learning.

🧠

The Approach

We use DINOv2 ViT-S/14 as a frozen feature extractor — a pretrained vision transformer from Meta that understands visual features without any domain-specific training. On top of it, we train a lightweight convolutional segmentation head that maps these features into 10 desert terrain classes.

📊

The Dataset

The Off-road Segmentation Dataset contains 960×540 desert terrain images with pixel-level annotations for 10 classes — from sky and trees to rocks and ground clutter.

⚡

Tech Stack

PyTorch DINOv2 Google Colab Chart.js HTML/CSS/JS