3D Coral Reef Reconstruction

3D Coral Reef Reconstruction

Atollkey is a hospitality business in Maldives where they provide a full immersive experience to cater to all clients, either relaxing or more adventurous activities. As part of the local businesses, the owner of Atollkey is set to contribute back to the community and would like to create educational opportunities for both locals and tourists while making an impact in conservation efforts. Focusing on the pressing issue of global warming and coastal health, coral reef is the focus for this project.

In this project, the focus is on the monitoring tasks where the cost is currently very high due to manual and tedious processes. The traditional monitoring process is called transect method where divers are required to mark out each coral reef individually. There is also another method called Photogrammetry which is a 3D reconstruction from various images from many different angles. It requires an elaborated planning of angles and coordinating these angles is not an easy task underwater. Hence, this project’s goal is to introduce AI powered automation for 3D reconstruction bundled as a user friendly application. Yielding increased accessibility and ease of use. Our objective is to introduce AI and computer vision as a more efficient substitute. 

Solution Approach

Our solution is a video to 3D construction application that simplifies the processes for the users, all the while leveraging the prowess of NeRF’s bleeding edge techniques and algorithms. With the reduction in usage complexity, our application will be accessible by a wide range of people coming from different backgrounds not limited to only experts in the field. 

The NeRF algorithm used is a product of relentless searching and testing multiple variations until finally settling upon the best fit for our purposes.

Project Objectives

Our project will encompass the following stages:

1.Research and Analysis: Conduct an in-depth review of existing NERF techniques and identify areas for improvement in the state-of-the-arts

2.Algorithm Optimization: Use python to optimize and test existing algorithms to reduce computational resource requirements and improve cost-effectiveness

3.Implementation and Testing: Implement the chosen NeRF method in a controlled environment and evaluate its performance.

4.User Interface Design: Design a user-friendly interface for both tourists and locals to engage with the monitoring process.

Model implementation


Implementing NeRF for underwater scenes involves unique challenges due to the distinct characteristics of underwater environments, such as color distortion, and varying levels of visibility. To adapt NeRF for such scenarios, several modifications and considerations are necessary. Firstly, the model must be able to account for the varying visibility levels in underwater scenes. Additionally, the training of the NeRF model needs adjustments to handle the color and light distortion caused by water. One such variation of NeRF called SeaThru-NeRF was engineered to account for such conditions.

Data collection

We have collected a great diversity of datasets for testing. The main tool is a gopro camera. The collected scenes from both underwater and on-land settings, and distinct subjects from each set together displaying a range of properties.

COLMAP software development

The COLMAP software varies between different NeRF variations. For DeblurNeRF, for example, it requires an external COLMAP GUI to generate files for NeRF. These include binary files for camera positions and such. However, Instant NGP has its own COLMAP system in its files. 


The rendered model from DeblurNeRF is quite blurry and low quality in general. This is due to the significant hardware limitations. The quality of the input images has been set to a lower quality version just for the program to output a result.

What we have done so far for the results were the same as summary of accomplishments, as we already mentioned as we already successfully installed and trained the NeRF such as De-Blur NeRF and Instance NGP NeRF (INGP) as it was our goals since the beginning of the project. 

Summary of Accomplishments

We’ve successfully installed and trained multiple variations of NeRF such as De-Blur NeRF and Instance NGP (INGP). The results have had varying quality levels depending on the size of the training data and the quality of the input data, but nonetheless, it was progress that can be built upon. Moreover, we already have a web application prototype for how our web application looks and the methodology to use it as well.application usable platform.

Future Directions

There are several methods to implement regression to potentially improve the experience of training NeRF and using Colmap. By incorporating gyro and accelerometer data to refine camera position estimates, and employing high-dimensional regression to predict optimal frame processing capacities based on specific computing resources, we have notably modeled a method to potentially enhance the efficiency and accuracy of these systems and the user experience by training them.

Sarutch Supaibulpipat
CMKL University
Matt Naganidhi
CMKL University
Natha Kunakorn
CMKL University
Settawut Chaithurdthum
CMKL University
Sivakorn Phromsiri
CMKL University
Dr. Permpoon Boonyarit