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.
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%.
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)