This repository provides the 21st-place solution for the Kaggle Airbus Ship Detection Challenge, offering a baseline for satellite image segmentation and detection. It's primarily aimed at Kaggle participants and researchers familiar with object detection and image segmentation tasks.

How It Works

The solution leverages the Detectron framework, a research platform for object detection and segmentation. It involves converting RLE (Run-Length Encoding) masks to COCO format for dataset compatibility, modifying Detectron's source code for the specific Airbus dataset, and configuring training parameters via YAML files. The approach utilizes Mask R-CNN, a popular model for instance segmentation.

Quick Start & Requirements

• Install/Run: Requires a Detectron environment. A Docker image (pascal1129/detectron:caffe2_cuda9_aliyun) is provided, or users can build from the official Detectron Dockerfile.
• Prerequisites: Caffe2, CUDA (version 9 specified in the provided Docker image), and Detectron.
• Setup: Modifying Detectron source files and configuring YAML files is necessary.

Highlighted Details

• Provides a 21st-place solution for a Kaggle competition.
• Includes scripts for RLE to COCO conversion.
• Offers guidance on modifying Detectron for the specific dataset.
• Includes scripts for generating submission CSV files.

Maintenance & Community

No specific information on maintenance, contributors, or community channels is provided in the README.

Licensing & Compatibility

The repository does not explicitly state a license. Compatibility with commercial or closed-source projects is not specified.

Limitations & Caveats

The provided Docker image is noted as "a little out of date." The setup requires modifying Detectron's source code, which can be fragile and difficult to maintain. Caffe2 installation is described as "troublesome."