Implementation of Sea-thru by Derya Akkaynak and Tali Treibitz
Forked from https://github.com/hainh/sea-thru
This fork's only aim is to standardize the dependencies and environment with Docker and Anaconda to make it so that anyone can use it without deep technical knowledge of python.
Additionally, in order to make this an all-in-one package, the Monodepth2 submodule has been cloned and commited directly into this repo, preserving the commit history and original authors.
- Run
docker compose up --build- This can take around 2 hours the first time; from then onwards you only need to run
docker compose upto get the container running and it'll boot up instantly.
- This can take around 2 hours the first time; from then onwards you only need to run
- Once the container is built and running, in a separate terminal run
docker exec -it seathru /bin/bash- Replace
seathruin that command if you've modified theSERVICEenviornment variable in.env
- Replace
- From within the container you can now run
python seathru-mono-e2e.py --image ${PATH_TO_IMAGE}- This root directory is mounted into the container each time you bring it up via
docker compose up, meaning any files you add on your host machine to this directory will also be available within the container.
- This root directory is mounted into the container each time you bring it up via
A recent advance in underwater imaging is the Sea-Thru method, which uses a physical model of light attenuation to reconstruct the colors in an underwater scene. This method utilizes a known range map to estimate backscatter and wideband attenuation coefficients. This range map is generated using structure-from-motion (SFM), which requires multiple underwater images from various perspectives and long processing time. In addition, SFM gives very accurate results, which are generally not required for this method. In this work, we implement and extend Sea-Thru to take advantage of convolutional monocular depth estimation methods, specifically the Monodepth2 network. We obtain satisfactory results with the lower-quality depth estimates with some color inconsistencies using only one image.