Underwater scenes are challenging for computer vision methods due to color degradation caused by the water column and detrimental lighting effects such as caustic caused by sunlight refracting on a wavy surface. These challenges impede widespread use of computer vision tools that could aid in ecological surveying of underwater environments or in industrial applications. Existing algorithms for alleviating caustics and descattering the image to recover colors are often impractical to implement due to the need for ground-truth training data, the necessity for successful alignment of an image within a 3D scene, or other assumptions that are infeasible in practice. In this paper, we propose a solution to tackle those problems in underwater computer vision: our method is based on two neural networks: CausticsNet, for single-image caustics removal, and BackscatterNet, for backscatter removal. Both neural networks are trained using an objective formulated with the aid of self-supervised monocular SLAM on a collection of underwater videos. Thus, our method does not requires any ground-truth color images or caustics labels, and corrects images in real-time. We experimentally demonstrate the fidelity of our caustics removal method, performing similarly to state-of-the-art supervised methods, and show that the color restoration and caustics removal lead to better downstream performance in Structure-from-Motion image keypoint matching than a wide range of methods.
We use self-supervised monocular SLAM neural networks to formulate a learning objective to estimate and remove the backscatter from underwater images.
Following the physical model of underwater image formation, the image arriving at the camera can be separated in the direct signal and the backscatter, which acts additively on the raw signal arriving at the camera sensor.
With some approximations, the backscatter of an image is governed by a handful of coefficients and the depth (z-axis). We train a neural network to predict the backscatter coefficients that minimize the re-projection error to a random overlapping image.
This learning objective allows to predict the backscatter of single unseen images and remove it. An overview of the training method is found below, for more details, consult the paper.
We use neural network to predict the expected reprojection error of an image with a random overlapping image, finding that this is a very effective way of removing caustics.
The reprojection error of two overlapping images is computed by using the estimated depths and poses by a monocular SLAM system. For a well-trained SLAM neural networks, the part of the reprojection error coming from mis-estimation of depths and poses is very small compared to the dominant noise caused by the caustics.
Training a neural network to predict the expected error for a single image without seeing the other image thus proves an effective way to remove caustics from images. A schematic overview is shown below, for more details, consult the paper.
Advances in computer vision and machine learning are the key to the next generation of coral monitoring tools. We are dedicated to creating scalable low-cost solutions that automate the tedious part of monitoring and allow us to see the change of coral reefs in the wake of climate change in unprecedented spatial and temporal resolution by dramatically increasing the amount of surveys that can be realized. The crucial components to this are semantic segmentation of benthic classes, and real-time 3D mapping of underwater environments on simple computing devices. The combination of these AI tools are embodied in the DeepReefMap project. The caustics and backscatter alleviation techniques presented in this research are a crucial cornerstone to being able to increase the quality of real-time 3D mapping techniques for reef environments.
A demo of the real-time 3D semantic mapping tool DeepReefMap
The Transnational Red Sea Center is a scientific research center created in 2019 at the Ecole Polytechnique fédérale in Lausanne (EPFL) with the official support of the Swiss Foreign Ministry. An independent and not-for-profit organization, the Center capitalizes on Switzerland’s neutrality, its longstanding tradition of promoting dialogue and its reputation for scientific excellence in order to bridge science and diplomacy for the future of coral reefs. In the Red Sea and beyond.
The Djibouti Mission / September 2022 from TRSC on Vimeo.
@inproceedings{sauder2024self,
title={Self-Supervised Underwater Caustics Removal and Descattering via Deep Monocular SLAM},
author={Jonathan Sauder and Devis Tuia},
booktitle={European Conference on Computer Vision},
year={2024},
organization={Springer}
}