MATL: Multi-Annotation Triplet Loss

This repository implements the Multi-Annotation Triplet Loss (MATL) method, as described in the paper accepted for Oral Presentation at IGARSS 2025.

Dataset:
AWIR dataset collected by Mississippi State Geosystems Research Institute.

Managed by Meilun Zhou through GatorSense.

---

Environment Setup

You can build the environment using either Conda or pip.

Using Conda (Recommended)

conda env create -f environment.yml
conda activate mmctl

Using pip

pip install -r requirements.txt

---

Python Files Overview

• model.py
  Contains the core model architectures, including:

  • Encoder and decoder networks for image and feature processing.
  • Functions for image reconstruction (with and without triplet loss).
  • Classifier architectures for 3-class and 9-class problems.
  • Utilities for building and summarizing Keras models.

• multi_annotation_triplet_loss.py
  Implements the triplet loss functions with online triplet mining:

  • Computes pairwise distances and valid triplet masks.
  • Defines the main triplet loss and a double-loss variant for multi-task learning (classification and bounding box aspect ratio).

• preprocessing.py
  Data preprocessing utilities for the AWIR dataset:

  • Extracts bounding boxes from annotation files.
  • Creates Pandas DataFrames from image folders and annotations.
  • Functions for plotting images with bounding boxes and filtering images by size.