Skip to content

federicatopazio/MARS_Image_Segmentation

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

3 Commits
 
 
 
 
 
 

Repository files navigation

Artificial Neural Networks and Deep Learning HW2 - AY 2024/2025

Challenge: Mars Terrain Segmentation

This challenge involved segmenting Mars terrain images into five classes: Background, Soil, Bedrock, Sand, and Big Rocks. We built a deep learning model from scratch, optimizing it for the Mean Intersection over Union (MeanIOU) metric.

Data & Augmentation

Dataset: 2,615 images, with 110 low-quality images removed. Maintaining the original class distribution provided the best results.

Key augmentations:

  • Horizontal & vertical flips
  • Exclusion of rotations & zoom to preserve features

Performance Boost: Flipping-based augmentation improved MeanIOU by 1%-6%.

Models & Training

Best Model: Multipath UNet with:

  • Multi-path encoders (Convolutional, Residual, Global Context, Multiscale)
  • Squeeze-and-Excitation bottleneck for feature fusion
  • Focal Loss with class balancing

Hyperparameter Tuning:

  • Optuna-based optimization for learning rate, batch size
  • Best optimizer: AdamW with weight decay

Final Results: MeanIOU 0.74 (Best Submission)

More Info

Refer to the report and notebooks.

Team

About

Image Segmentation Challenge on Mars terrain pictures

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published