Airbus Ship Detection – Semantic Segmentation Project

This repository contains code and models for ship detection in satellite images using deep learning-based semantic segmentation. It leverages a cleaned and preprocessed subset of the Airbus Ship Detection dataset. After preprocessing and cleaning, a subset of 12,788 images is obtained, which is then cached into local storage for faster lookup when training the model.

---

Generating the ready-to-use dataset

Download the subset of 12,788 cleaned and preprocessed images from Google Drive and place them in some folder.

Note: In case we wanted to start to train the models, we would need to first generate the cached images and masks, by also downloading the cleaned masks_subset.csv file from Google Drive (or project folder), updating the 4 path variables at the beginning of the file 'preprocessing/build_cache.py', and running it as:

python preprocessing/build_cache.py

Using inference.ipynb notebook on test set

Note that, currently, this notebook already has some test set image examples saved in the folder images within the current directory. However, one can also make it work for any image on the test set by doing the following:

1. Download the test set of 15,606 images from Google Drive
2. In cell 4 within the notebook, set the variable TEST_DIR to the path where these images were saved, and uncomment the lines stated in the notebook.

Our best model is the finetuned_model_512.tflite and exists in the repo, you can also find other models in our Google Drive, if you wish to use them for inference, you will need to convert them using lighter_model.py, this will convert the .keras to .tflite, which makes the model faster for inference and adapted to our script.

---

Repository Structure

Root Files

• README.md – Project documentation and setup instructions
• requirements.txt – Python package dependencies if you use pip
• environment.yaml – Python package dependencies for Mac
• environment_windows.yaml – Python package dependencies for Windows
• masks_subset.csv – CSV file containing ship mask annotations for the cleaned subset
• inference.ipynb – Jupyter notebook for running inference on test images using trained models

Directories

preprocessing/

Contains scripts and notebooks for data preprocessing and exploration:

• build_cache.py – Generates cached images and masks for faster training
• helpers.py – Utility functions for preprocessing
• dataprocessing.ipynb – Data processing exploration and workflow
• image_creation.ipynb – Image preparation and augmentation
• eda-of-the-dataset.ipynb – Exploratory Data Analysis
• check_images.ipynb – Image validation and quality checks
• old_explorations/ – Archive of earlier exploration notebooks

trainings/

Core training scripts for different model architectures:

• FINAL_MODEL.py – Final optimized model training script
• ResNet34.py – ResNet34-based semantic segmentation model
• ResNet101.py – ResNet101-based semantic segmentation model
• UNet.py – U-Net based semantic segmentation model
• npy_generator.py – Data generator for training from numpy arrays

postprocessing/

Post-training optimization and inference scripts:

• postprocessing.py – Post-processing pipeline for model predictions
• submission_parallel.py – Parallel inference for generating submission files
• submission_512.py – Submission script optimized for 512-pixel images
• finetuning.py – Fine-tuning script for model improvements
• lighter_model.py – Model optimization for reduced size/inference time
• debug_model.py – Debugging utilities for model analysis
• debug_model_big.py – Extended debugging utilities

models/

Trained model files:

• final_model.tflite – Final optimized TensorFlow Lite model
• finetuned_model_512.tflite – Fine-tuned model for 512-pixel inputs

results/

Training results and visualizations:

• initial_train/ – Results from initial model training
  • training_metrics.csv – Initial training metrics
• finetuning/ – Results from fine-tuning experiments
  • training_metrics.csv – Fine-tuning metrics
• visual_debug_50/ – Visual debugging outputs (50 samples)
• visual_debug_old_50/ – Archive of older visual debugging outputs

final_submissions/

Final submission files for competitions/evaluations:

• resnet34.csv – ResNet34 predictions
• resnet34FN.csv – ResNet34 with false negative handling
• resnet34_FN+filtered.csv – ResNet34 with filtered false negatives
• resnet101.csv – ResNet101 predictions
• unet.csv – U-Net predictions

cluster/

Scripts for distributed/cluster computing:

• run.sh – Main cluster execution script
• run2.sh – Alternative cluster execution variant
• run3.sh – Another cluster execution variant

logs/

Logging utilities:

• log_processing.py – Script for processing and analyzing logs

images/

Test images for inference demonstrations

---