Skip to content

README.md

Latest commit

 

History

History
96 lines (65 loc) · 3.19 KB

README.md

File metadata and controls

96 lines (65 loc) · 3.19 KB
Raw

Semantic Segmentation Evaluation

Example script for training semantic segmentation models using pretrained DeepAndes backbone with a simple linear segmentation head. Also supports multiple baseline models including MAE, MoCoV2, SatMAE, and training from scratch.

Overview

This repository contains code for evaluating semantic segmentation performance on 8-band satellite imagery. The training script performs 5-fold cross-validation and tracks metrics using Weights & Biases.

Dataset Format

The dataset should be organized as follows:

dataset_folder/
├── images/
│   ├── image_1.npy
│   ├── image_2.npy
│   └── ...
└── masks/
    ├── image_1.npy
    ├── image_2.npy
    └── ...
  • Images: NumPy arrays (.npy format) containing 8-band satellite imagery
  • Masks: NumPy arrays (.npy format) containing binary segmentation masks
  • Example dataset is provided in active_corrals_data

Configuration Files

The configs contains example YAML configuration files for different model backbones on activate corral segmentation task.

The model_name in config YAML file supports the following backbone options:

  • deepandes — our ViT-L model from DINOv2
  • mae — Masked Autoencoder
  • mocov2 — Momentum Contrast v2
  • satmae — A Satellite MAE baseline
  • scratch — randomly initialized ViT-L (no pre-training)

The MoCoV2 pre-trained weight (moco_v2_200ep_pretrain.pth.tar) can be downloaded from offical github download here.

Usage

Training

Train a model using a configuration file:

python main_binary_experiment.py --config /path/to/config.yaml

Example: Active Corral Segmentation

Train using the DeepAndes (FM3M, pretrained 3 million) configuration:

python main_binary_experiment.py --config configs/corrals_active_FM3M.yaml

Configuration Setup

Before training, edit the configuration YAML file to set the correct paths:

  1. Data paths: Update image_folder and mask_folder to point to your dataset
  2. Output directory: Set output_dir to where you want to save model checkpoints and visualizations
  3. Weights & Biases: Update project and name for experiment tracking
  4. Pretrained weights: If using a pretrained model, specify the path in pretrained_weights (if necessary)

Training Details

  • Cross-validation: 5-fold cross-validation is performed automatically
  • Evaluation metrics: Dice Score and Mean IoU
  • Visualizations: Predictions are saved every 25 epochs (if output_dir is specified)
  • Backbone freezing: Set train_backbone: False to freeze the backbone and only train the segmentation head

Outputs

When output_dir is specified, the script saves:

  • Model checkpoints at epoch intervals
  • Validation visualizations (RGB image, prediction overlay, ground truth)
  • Prediction masks (NumPy format) for each validation sample

Output structure:

output_dir/
└── fold{fold_number}_{epoch}ep/
    ├── image_0.jpg
    ├── image_0.npy
    ├── image_1.jpg
    └── ...