Coral me is a collaborative open source system dedicated to underwater coral reef image annotation.
Global warming and local anthropogenic stressors are causing severe stress to coral reefs across the world. To take appropriate action decision makers need accurate data over large spatio-temporal scales.
The speed of data collection has increased tremendously in recent years, and as a result, millions of images are collected each years across the world. The analysis process remains painfully slow as manual inspection of each photo is often required by trained experts.
This project sets for itself the following goals
CoralMe is more than a bundle of state-of-the-art code. It aims to be a platform alloying computer vision researchers to easily experiment, improve and contribute to coral-reef monitoring all in the comfort of their usual environment. It gives researchers worldwide, the very marine biologists documenting and investigating the health of marine life, unprecedented access to the most recent advances in computer vision without having to transition to new sets of tools. It finally offers the opportunity for anyone to actively participate in monitoring the health of coral reefs across the oceans.
There are many annotation platforms and software available and it is difficult and costly for a group to transition from one tool to another. A new software is simply not the solution. CoralMe allows you to keep your tools that have been finely tuned over the years, while still benefiniting from new technology at a low cost.
The Smart Region selector is a segmentation tool that allows the user to quickly select a region with a distinct texture. Developed internally to quickly establish ground truths over large datasets, we’ve now made it available in CoralMe. We hope this tool will help others!
Superpixels are a popular approach to segmentation, having been used succesfully in other projects (i.e. photoQuad, Seascape). It provides a viable and easily interpretable alternative to handfree region selection for full image annotation.
Manual or automated segmentation and annotation will often leave certain pixels unannotated. These often lie near boundaries where two coral colonies meet. This refinement tool allows the user to snap all unassigned pixels to their nearest and most similar region using unsupervised Gaussian Mixture Modeling.
Machine learning is in constant evolution with new ideas and approaches helping us squeeze more accurate measurements. CoralMe’s deep learning is based on the easy-to-use convolutional neural networks library MatConvNet. Deep learning models are used as a powerful transfer learning feature extraction methods.
Manual annotation is tedious, frustrating and prone to error... And yet very much necessary. We understand that. Much of our recent development has been towards facilitating this process and making sure it fulfills the needs of both on-the-ground researcher and Machine Learning practicioner to complete a self-enhancing and functionning loop of collaboration.
This short video aims to show what a handful of passionate individual can do.